Graph attention network-enhanced multi-agent reinforcement learning for dynamic interception task allocation in counter-drone defense Authors: Dianbo Jia, Ganliang Wang, Hongfei Bu, Junwei Wang, Zhengkai Wang Published: 2026 ABSTRACT In cooperative counter-drone defense, dynamic interception task allocation presents a fundamental challenge: targets may appear, vanish, or shift trajectory unpredictably throughout an engagement. Current multi-agent reinforcement learning approaches generally lack explicit mechanisms to capture the changing correspondence between interceptors and targets as the operational topology evolves. We propose DT-GAT-MARL, a hierarchical framework pairing a Dynamic-Topology Graph Attention Network (DT-GAT) at the strategic allocation level with Multi-Agent Proximal Policy Optimization (MAPPO)-based maneuver control at the tactical level. Three design choices define DT-GAT: a masking mechanism that handles node additions and removals without rebuilding the graph, learnable edge-feature biases that embed spatial reachability and interception urgency directly into attention computation, and a Gumbel-Softmax allocation head that enables differentiable discrete assignment while preserving end-to-end gradient flow. A dual-frequency architecture separates the allocation and maneuvering timescales, which helps suppress assignment oscillations. We evaluate the framework across balanced, numerically disadvantaged, and dynamic intrusion scenarios ranging from 4v4 to 12v12. In the dynamic intrusion setting, DT-GAT-MARL exceeds the strongest baseline by 10.3 percentage points, reaches an 87.3% effective reallocation rate, and keeps oscillation at just 9.6%. Ablation results confirm that edge-feature bias is the single most critical component. Introduction Low-cost commercial drones have proliferated rapidly in recent years, posing growing threats to critical infrastructure, public safety, and military installations—making counter-drone defense an increasingly pressing priority1. Recent studies have further explored deep reinforcement learning for autonomous combat planning in adversarial UAV environments2. Meanwhile, Tian et al.3 revealed that deep-learning-based UAV systems are susceptible to adversarial attacks on both navigation and control modules, further underscoring the security challenges that hostile conditions pose to autonomous drone platforms and reinforcing the urgency of robust defensive coordination. Compared with conventional manned air combat, countering drone incursions introduces distinct difficulties: target numbers shift dynamically, threat arrival sequences are uncertain, and swarm-scale coordination effects emerge4. Together, these factors make real-time task allocation the central decision-making bottleneck5. For cooperative interception to succeed, defender clusters must tightly couple target assignment with maneuver control amid continuously evolving tactical conditions, demanding both responsiveness and adaptability. Classical combinatorial optimization has received considerable attention in this area. The consensus-based auction algorithm of Choi et al6. achieves robust allocation through distributed bidding and consensus propagation, though its dependence on iterative convergence constrains applicability when targets appear or vanish in real time. Skaltsis et al.7, in their broad survey, observed that the computational cost of traditional optimization scales sharply with agent count, creating severe real-time bottlenecks at larger scales. Game-theoretic formulations offer another avenue—Ren et al.8 applied incomplete-information dynamic games to multi-UAV maneuvering decisions and showed the utility of equilibrium strategies for adversarial reasoning. Such methods, however, assume a finite discrete strategy space and quickly become intractable in high-dimensional continuous action settings. Deep reinforcement learning (DRL) has introduced a data-driven, end-to-end paradigm for multi-agent coordination. On the value decomposition side, Sunehag et al9. proposed VDN, which additively decomposes team rewards across agents—a foundational step toward decentralized cooperative learning. In the policy gradient direction, MADDPG10 established centralized-training-decentralized-execution (CTDE) as a key baseline for continuous multi-agent control, and HAPPO11 later added trust-region constraints guaranteeing monotonic policy improvement across heterogeneous agents. A shared limitation of these methods, however, is that cooperative relationships stay implicitly buried in network parameters; none explicitly model the dynamic correspondence between interceptors and targets. Graph neural networks (GNNs) offer a natural complement. Kipf et al.12 demonstrated that graph networks can infer latent relational structures from observational data, while Liu et al.13 showed through G2ANet that graph attention mechanisms automatically extract critical interaction patterns in multi-agent games, substantially improving coordination. More recently, graph attention mechanisms have been applied to other cooperative decision-making domains: for instance, Zhang14 employed a graph attention-based heterogeneous multi-agent framework for adaptive portfolio optimization, demonstrating that attention-driven inter-agent communication generalizes beyond robotic control to strategic resource allocation. Even with this progress, existing GNN-based methods remain insufficient for counter-drone interception in three respects. First, most rely on static or predefined graph topologies that cannot gracefully handle node additions or removals—whether from successful interceptions or newly detected intruders—and rebuilding the graph discards historical information at unnecessary computational cost. Second, standard graph attention derives aggregation weights from node features alone, overlooking geometric cues carried by edges; spatial reachability and interception urgency, for instance, would allow interceptors to modulate attention based on relative tactical posture. Third, the inherent timescale gap between strategic-level discrete assignment and tactical-level continuous maneuver control has lacked an explicit coordination mechanism in prior work. To address these gaps, we propose DT-GAT-MARL, a multi-agent reinforcement learning framework built around a dynamic-topology graph attention network and organized as a two-tier decision architecture linking high-level allocation with low-level maneuvering. Our contributions are as follows. We design the DT-GAT module to support node-level additions and removals via masking, incorporating learnable edge-feature biases into attention computation to encode spatial relationships, and employing a Gumbel-Softmax allocation head for differentiable discrete assignment. We construct a dual-frequency, loosely coupled hierarchical framework where the high-level DT-GAT produces allocation plans at low frequency and the low-level MAPPO executes maneuver control every timestep, with gradient connectivity maintained through Gumbel-Softmax and a pretrain-then-jointly-finetune training strategy. We conduct comprehensive experiments spanning balanced, disadvantaged, and dynamic-intrusion scenarios—from 4v4 to 12v12 scalability tests and seven ablation configurations—systematically quantifying each module’s contribution and validating real-time reallocation under dynamic intrusions. Problem Formulation and modeling Scenario description and kinematics As illustrated in Fig. 1, we consider a two-dimensional counter-drone interception scenario. Interceptors (defenders) deploy along the forward edge of a protected zone, while intruding drones approach from various directions. Every UAV follows a planar three-variable kinematic model, widely adopted in multi-UAV air combat research15. The state equations for the i-th UAV at time t are:Fig. 1Anti-UAV interception scenario in 2D plane.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\dot{x}}_{i}={v}_{i}\mathrm{cos}{\psi}_{i},{\dot{y}}_{i}={v}_{i}\mathrm{sin}{\psi}_{i},{\dot{\psi }}_{i}={\omega}_{i}$$\end{document} Anti-UAV interception scenario in 2D plane. We adopt the planar kinematic model as a standard baseline formulation that is widely used in multi-UAV interception studies8,15 and permits direct comparison with existing methods. Intruders follow constant-velocity straight-line trajectories toward the protected zone, representing a worst-case penetration strategy in which the adversary prioritizes reaching the target over evasion. While this assumption does not capture evasive or cooperative intruder behavior, it provides a well-controlled testbed for isolating the contribution of the allocation mechanism–-the primary focus of this work. Here \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left({x}_{i},{y}_{i}\right)$$\end{document} denotes position, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{i}$$\end{document} the flight speed, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\psi}_{i}$$\end{document} the heading angle. Speed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{i}$$\end{document} and turn rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\omega}_{i}$$\end{document} serve directly as control inputs, subject to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{i}\in [{v}_{\mathrm{m}\mathrm{i}\mathrm{n}},{v}_{\mathrm{m}\mathrm{a}\mathrm{x}}]$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mid {\omega}_{i}\mid \le {\omega}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document}. Interception succeeds when the distance between defender i and intruder j satisfies \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}_{ij}\le {d}_{cap}$$\end{document}, upon which intruder j is removed and defender i reverts to an idle state available for reassignment. An intruder reaching the defense zone boundary counts as a breach. The episode terminates once all intruders have been either intercepted or breached; any intruder still active at the maximum step limit is treated as a breach. Table 1 summarizes kinematic parameters and node feature definitions.Table 1Kinematic parameters and node feature definitions for defenders and intruders.CategorySymbolMeaningDefenderIntruderKinematicsvFlight speed (m/s)[180, 250][120, 200]\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\omega}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document}Max turn rate (rad/s)0.150.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}_{cap}$$\end{document}Capture distance (m)50—Δ tSimulation timestep (s)0.10.1Node featuresx,yPosition (m)✓✓vCurrent speed (m/s)✓✓ψHeading angle (rad)✓✓\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{status}$$\end{document}Status flagIdle/TrackingActive/Intercepted Table 1Kinematic parameters and node feature definitions for defenders and intruders.CategorySymbolMeaningDefenderIntruderKinematicsvFlight speed (m/s)[180, 250][120, 200]\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\omega}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document}Max turn rate (rad/s)0.150.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}_{cap}$$\end{document}Capture distance (m)50—Δ tSimulation timestep (s)0.10.1Node featuresx,yPosition (m)✓✓vCurrent speed (m/s)✓✓ψHeading angle (rad)✓✓\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{status}$$\end{document}Status flagIdle/TrackingActive/Intercepted Kinematic parameters and node feature definitions for defenders and intruders. DEC-POMDP formulation under CTDE We model the multi-interceptor coordination problem as a decentralized partially observable Markov decision process (DEC-POMDP)16,17, specified by the tuple \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\langle \mathcal{N},\mathcal{S},\{{\mathcal{O}}_{i}\},\{{\mathcal{A}}_{i}^{h},{\mathcal{A}}_{i}^{l}\},P,R,\gamma \rangle$$\end{document}. Here \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{N}=\{1,\dots ,{N}_{d}\}$$\end{document} is the set of interceptors; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{S}$$\end{document} is the global state space encompassing all UAV states and assignment relations; and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{O}}_{i}$$\end{document} is the local observation of interceptor i, limited to its own state and targets within communication range. Concretely, the local observation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{o}}_{i}\in {\mathbb{R}}^{{d}_{o}}$$\end{document} of interceptor i comprises its own kinematic state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left({x}_{i},{y}_{i},{v}_{i},{\psi}_{i},{s}_{status,i}\right)$$\end{document} and, for each intruder j satisfying \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\parallel {\mathbf{p}}_{i}-{\mathbf{p}}_{j}\parallel \le {R}_{comm}$$\end{document}, the relative state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(\Delta {x}_{ij},\Delta {y}_{ij},{v}_{j},{\psi}_{j},{s}_{status,j}\right)$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{comm}=800$$\end{document} m is the communication range. Intruders outside this range are zero-padded. The observation dimension is therefore \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}_{o}=5+5\times {N}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document} is the maximum intruder capacity. The global state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{S}$$\end{document} aggregates the kinematic states of all UAVs together with the current allocation matrix. The action space follows a hierarchical design: the high-level action \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{A}}_{i}^{h}$$\end{document} is a discrete target assignment decision updated every K steps or upon event triggers, while the low-level action \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{A}}_{i}^{l}$$\end{document} consists of continuous velocity-level commands \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left({v}_{i},{\omega}_{i}\right)$$\end{document} consistent with the kinematic model, executed at every step. Only these two decision tiers—allocation and maneuvering—are present in the framework. P denotes the transition function, R a shared reward, and γ the discount factor. Training follows the CTDE paradigm18: the Critic network accesses the full state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{S}$$\end{document} for value estimation, whereas each interceptor acts solely on its local observation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{O}}_{i}$$\end{document} and the high-level assignment directive during execution. Dynamic bipartite graph representation To explicitly capture the evolving interactions between interceptors and intruders, we represent the tactical situation as a dynamic bipartite graph \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{G}}_{t}=({\mathcal{V}}_{t}^{d}\cup {\mathcal{V}}_{t}^{u},{\mathcal{E}}_{t})$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{V}}_{t}^{d}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{V}}_{t}^{u}$$\end{document} are the active defender and intruder node sets at time t, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{E}}_{t}$$\end{document} is a fully connected edge set. Because the two node types have different raw feature dimensions (see Table 1), type-specific linear projections map them into a shared d-dimensional embedding space:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z}_{i}^{d}={W}^{d}{h}_{i}^{d}+{b}^{d},{z}_{j}^{u}={W}^{u}{h}_{j}^{u}+{b}^{u}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h}_{i}^{d}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h}_{j}^{u}$$\end{document} are raw node features, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}^{d}\in {\mathrm{R}}^{d\times {d}_{1}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}^{u}\in {\mathrm{R}}^{d\times {d}_{2}}$$\end{document} are learnable projection matrices, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${b}^{d}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${b}^{u}$$\end{document} are bias vectors. Each edge \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(i,j)\in {\mathcal{E}}_{t}$$\end{document} carries an edge feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f}_{ij}=[\Delta {r}_{ij},\hspace{0.17em}\Delta {\theta}_{ij},\hspace{0.17em}{\widehat{t}}_{ij}]$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {r}_{ij}=\parallel {p}_{i}^{d}-{p}_{j}^{u}\parallel$$\end{document} is the Euclidean distance, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {\theta}_{ij}=atan2({y}_{j}^{u}-{y}_{i}^{d},\hspace{0.17em}{x}_{j}^{u}-{x}_{i}^{d})-{\psi}_{i}$$\end{document} is the relative bearing angle between the interceptor’s heading and the target’s direction, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{t}}_{ij}=\Delta {r}_{ij}/({v}_{i}-{v}_{j}\mathrm{c}\mathrm{o}\mathrm{s}\Delta {\theta}_{ij})$$\end{document} is an estimated intercept time based on constant-speed pursuit geometry. When the denominator is non-positive—meaning the interceptor cannot close distance along the target’s bearing—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{t}}_{ij}$$\end{document} is clamped to a preset upper bound \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document}. Topology changes are handled by two binary mask vectors: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{m}}_{t}^{u}\in \{\mathrm{0,1}{\}}^{{N}_{\mathrm{m}\mathrm{a}\mathrm{x}}}$$\end{document} for intruders and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{m}}_{t}^{d}\in \{\mathrm{0,1}{\}}^{{N}_{\mathrm{m}\mathrm{a}\mathrm{x}}}$$\end{document} for defenders, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{\mathrm{m}\mathrm{a}\mathrm{x}}$$\end{document} are the respective maximum node capacities. When an intruder is intercepted, its mask bit \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${m}_{j}^{u}$$\end{document} is set to 0; when a new intruder is detected, the corresponding bit is flipped to 1 with zero-initialized features. On the defender side, a destroyed or disabled interceptor (not modeled in the current scenarios but supported by the architecture) would similarly have \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${m}_{i}^{d}$$\end{document} set to 0. In the attention computation (Eqs. 4-5), masking is applied to the intruder dimension j because the bipartite attention flows from defenders to intruders; defender-side masking would take effect if defenders could be removed during an episode. This avoids rebuilding the graph or modifying network parameters: newly added nodes have their mask bit flipped from 0 to 1 with zero-initialized features, while removed nodes are simply masked out, and subsequent attention computations automatically skip them. Methodology Overall architecture DT-GAT-MARL adopts a dual-frequency, loosely coupled hierarchical architecture with two decision tiers having separate network parameters and optimizers. The high-level tier is the DT-GAT module, which outputs interceptor-to-target assignments at low frequency. The low-level tier is a MAPPO-based maneuver policy network producing velocity-level control commands every timestep. Information flows between tiers through the Gumbel-Softmax allocation head: assignment probabilities produce a weighted sum over intruder embeddings, yielding a target feature embedding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widetilde{z}}_{i}$$\end{document} that serves as auxiliary input to the low-level Actor. Specifically, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widetilde{z}}_{i}={\sum}_{j:{m}_{j}=1}{\pi}_{ij}{z}_{j}^{u}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\pi}_{ij}$$\end{document} is the Gumbel-Softmax output from Eq. (7). Since Gumbel-Softmax is differentiable, gradients from the low-level reward signal propagate back through to the high-level DT-GAT parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta}_{h}$$\end{document}, enabling end-to-end joint optimization. High-level allocation triggers on whichever comes first: periodic invocation (every K steps) or an event trigger (target intercepted, new target detected, or interceptor released). Between triggers, allocation remains frozen and the low-level policy continues maneuvering under the standing assignment. This lets the strategic layer update at a slower cadence while the tactical layer operates at full timestep resolution, explicitly reconciling their inherent timescale mismatch. The allocation follows a soft many-to-one scheme: each interceptor is assigned to exactly one target at a time, but multiple interceptors may independently select the same intruder if the attention mechanism deems it the highest-priority threat. No explicit capacity constraint is enforced; instead, the edge-feature bias naturally diversifies assignments because interceptors farther from an already-targeted intruder receive lower attention scores for it. In disadvantaged scenarios where intruders outnumber interceptors, each interceptor pursues one target at a time and, upon successful interception, is event-triggered to reallocate toward the next highest-priority active intruder, enabling sequential multi-target engagement. DT-GAT module The DT-GAT module is the core component of the proposed method. It augments the standard graph attention mechanism with three design choices: edge-feature-enhanced attention, a masking mechanism, and a Gumbel-Softmax allocation head. Edge-feature-enhanced attention In standard GAT19, attention logits depend solely on node features and thus cannot exploit the spatial geometry encoded in edges. Unlike GraphSAGE20, which relies on fixed aggregation functions, we introduce a learnable edge-feature bias into the attention computation, enabling data-driven adaptive aggregation. For attention head \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k\in \{1,\dots ,H\}$$\end{document}, the attention logit from interceptor i to intruder j is:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$e_{{ij}}^{{\left( k \right)}} = {\mathrm{LeakyReLU}}\,\left( {{\mathrm{a}}_{k}^{{ \top }} \left[ {{\mathrm{W}}_{k} {\mathrm{z}}_{i}^{d}\mid \,\parallel\,\mid {\mathrm{W}}_{k} {\mathrm{z}}_{j}^{u} } \right]{\text{ + }}\beta _{k}^{{ \top }} {\mathrm{f}}_{{ij}} } \right)$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{W}}_{k}\in {\mathrm{R}}^{{d}^{^{\prime}}\times d}$$\end{document} is the linear projection for head k, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{a}}_{k}\in {\mathrm{R}}^{2{d}^{^{\prime}}}$$\end{document} is the attention parameter vector, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\beta}_{k}\in {\mathrm{R}}^{{d}_{f}}$$\end{document} is the edge-feature bias parameter. The term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\beta}_{k}^{{\top}}{\mathrm{f}}_{ij}$$\end{document} has a clear physical interpretation: relative distance and bearing encoded in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{f}}_{ij}$$\end{document} allow attention weights to incorporate spatial reachability21, while the estimated intercept time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{t}}_{ij}$$\end{document} injects an urgency prior that steers interceptors toward more threatening and reachable targets. The idea of incorporating edge features into graph neural networks has been explored in prior work (e.g., Gong and Cheng21 proposed edge-feature-enhanced GNNs for general graph classification). Our contribution is not the concept of edge-feature bias itself, but its specific instantiation for multi-agent interception: we encode relative distance, bearing, and estimated intercept time as the edge feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${f}_{ij}$$\end{document}, giving the bias a concrete physical interpretation tied to spatial reachability and interception urgency. Combined with the masking mechanism and Gumbel-Softmax head—neither of which appear together in prior GNN-based MARL work for task allocation—the integration addresses the specific challenge of differentiable dynamic assignment under changing topologies. Masking mechanism To handle dynamic node additions and removals, the mask vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{m}}_{t}$$\end{document} suppresses invalid nodes before softmax normalization:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widetilde{e}}_{ij}^{\left(k\right)}=\left\{\begin{array}{cc}{e}_{ij}^{\left(k\right)},& \text{if }{m}_{j}^{u}=1\\ -\infty ,& \text{if }{m}_{j}^{u}=0\end{array}\right.$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\alpha}_{ij}^{\left(k\right)}=\frac{\mathrm{e}\mathrm{x}\mathrm{p}({\widetilde{e}}_{ij}^{\left(k\right)})}{\sum_{{j}^{^{\prime}}:{m}_{{j}^{^{\prime}}}^{u}=1}\mathrm{e}\mathrm{x}\mathrm{p}({\widetilde{e}}_{i{j}^{^{\prime}}}^{\left(k\right)})}$$\end{document} Adding a node requires only flipping the corresponding mask bit to 1 and zero-initializing its features; removing one sets the bit to 0. No structural or parametric changes are needed, preserving computational continuity under frequent topology changes. The multi-head aggregation output is:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{z}}_{i}^{{\mathrm{'}}} = \parallel _{{k = 1}}^{H} \mid\sigma \left( {\mathop \sum \limits_{{j:\,m_{j} = 1}} \alpha _{{ij}}^{{\left( k \right)}} \,{\mathrm{W}}_{k} {\mathrm{z}}_{j}^{u} } \right)$$\end{document}where σ denotes the ELU activation and ∥ is concatenation, producing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{z}}_{i}^{^{\prime}}\in {\mathrm{R}}^{H{d}^{^{\prime}}}$$\end{document}. Gumbel-Softmax allocation head The aggregated embedding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z}_{i}^{^{\prime}}$$\end{document} is projected into the same dimensionality as the intruder embeddings via \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{z}}_{i}={W}_{l}{z}_{i}^{^{\prime}}+{b}_{l}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{l}\in {\mathbb{R}}^{d\times H{d}^{^{\prime}}}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${b}_{l}\in {\mathbb{R}}^{d}$$\end{document}. The allocation logit for interceptor i toward intruder j is then computed as a bilinear score: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{ij}={\widehat{z}}_{i}^{{\top}}{W}_{b}{z}_{j}^{u}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{b}\in {\mathbb{R}}^{d\times d}$$\end{document} is a learnable bilinear weight matrix. This formulation ensures that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{ij}$$\end{document} explicitly depends on both the interceptor’s aggregated representation and each individual intruder’s embedding, addressing the need for target-specific scoring. Since task assignment is inherently a discrete choice. , a naive argmax would block gradient flow. We therefore apply Gumbel-Softmax reparameterization22,23:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi _{{ij}} = \frac{{{\mathrm{exp}}\,\left( {\left( {l_{{ij}} + g_{j} } \right)/\tau } \right)}}{{\mathop \sum \nolimits_{{j^{\prime}:\,m_{{j^{\prime}}} = 1}} {\mathrm{exp}}\,\left( {\left( {l_{{ij^{\prime}}} + g_{{j^{\prime}}} } \right)/\tau } \right)}}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${g}_{j}=-\mathrm{l}\mathrm{o}\mathrm{g}(-\mathrm{l}\mathrm{o}\mathrm{g}({U}_{j}))$$\end{document} is Gumbel noise with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${U}_{j}\sim Uniform(\mathrm{0,1})$$\end{document}, and τ is the temperature parameter. During training, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\pi}_{ij}$$\end{document} forms a continuous approximation to a one-hot vector, ensuring gradient propagation; at inference, a hard argmax yields deterministic assignments. The temperature τ is linearly annealed from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{0}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{\mathrm{m}\mathrm{i}\mathrm{n}}$$\end{document} during the second training stage—high temperatures early on encourage exploration of diverse allocations, while low temperatures later drive convergence toward near-deterministic policies. Maneuver policy and reward design The low-level maneuver policy uses an Actor-Critic architecture. The Actor network incorporates a GRU unit for temporal memory and takes as input the concatenation of interceptor i’s local observation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{o}}_{i}$$\end{document} and its assigned target embedding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbf{z}}_{target,i}$$\end{document}, outputting control commands \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left({v}_{i},{\omega}_{i}\right)$$\end{document}. The Critic, active only during training, estimates the state value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$V(\mathcal{S})$$\end{document} from the global state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{S}$$\end{document}, consistent with the CTDE paradigm18. Optimization follows MAPPO24, whose parameter sharing and value normalization techniques have demonstrated strong sample efficiency and training stability in cooperative multi-agent settings. The reward design prioritizes task-level signals, deliberately avoiding hand-crafted cooperative rules. Global rewards—interception success \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{cap}$$\end{document} and target breach \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{pen}$$\end{document}—are shared across both tiers so that allocation decisions are optimized directly toward task objectives. Shaping rewards act only on the low level to guide basic maneuvering behavior and include distance reduction, heading alignment, out-of-bounds penalties, and time efficiency signals. Additionally, a switching cost \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{switch}$$\end{document} discourages allocation oscillations, ensuring interceptors retain sufficient execution windows after reassignment. Table 2 details each reward component.Table 2Reward function components.TierComponentExpressionValueIntentGlobal (shared)Interception success\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{cap}$$\end{document} + 20Core task signalTarget breach\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{pen}$$\end{document} − 50Heavy penalty for defense failureLow-level shapingDistance reduction\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{dist}={\lambda}_{d}(\Delta {r}_{ij}^{t-1}-\Delta {r}_{ij}^{t})$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda}_{d}=0.5$$\end{document}Guide approachHeading alignment\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{head}={\lambda}_{h}\mathrm{c}\mathrm{o}\mathrm{s}({\psi}_{i}-{\theta}_{ij})$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda}_{h}=0.3$$\end{document}Reduce detoursOut-of-bounds\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{oob}$$\end{document} − 5Constrain operating areaTime penalty\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{time}$$\end{document} − 0.1/stepEncourage efficiencySwitching cost\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{switch}$$\end{document} − 2Suppress allocation oscillation Table 2Reward function components.TierComponentExpressionValueIntentGlobal (shared)Interception success\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{cap}$$\end{document} + 20Core task signalTarget breach\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{pen}$$\end{document} − 50Heavy penalty for defense failureLow-level shapingDistance reduction\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{dist}={\lambda}_{d}(\Delta {r}_{ij}^{t-1}-\Delta {r}_{ij}^{t})$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda}_{d}=0.5$$\end{document}Guide approachHeading alignment\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{head}={\lambda}_{h}\mathrm{c}\mathrm{o}\mathrm{s}({\psi}_{i}-{\theta}_{ij})$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda}_{h}=0.3$$\end{document}Reduce detoursOut-of-bounds\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{oob}$$\end{document} − 5Constrain operating areaTime penalty\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{time}$$\end{document} − 0.1/stepEncourage efficiencySwitching cost\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{switch}$$\end{document} − 2Suppress allocation oscillation Reward function components. A potential concern with shared global rewards is the credit assignment problem, where individual agents may adopt free-riding behavior. Three design features mitigate this risk in DT-GAT-MARL. First, the low-level shaping rewards (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{dist}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{head}$$\end{document}) are computed per interceptor based on its individual trajectory toward the assigned target, providing agent-specific learning signals. Second, the DT-GAT allocation mechanism assigns each interceptor a distinct target, creating implicit task specialization that ties each agent’s maneuvering behavior to an identifiable subtask. Third, MAPPO’s parameter sharing and value normalization have been empirically shown to alleviate lazy-agent tendencies in cooperative settings24. Empirically, we monitored each interceptor’s individual interception count across 250 evaluation episodes in the 8v8 scenario: the coefficient of variation (CV) across agents was 0.12, indicating approximately balanced workload distribution rather than concentration on a few active agents. The reward weights (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda}_{d}=0.5$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda}_{h}=0.3$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{cap}=+20$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{pen}=-50$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{switch}=-2$$\end{document}) were determined through a systematic grid search over 36 configurations in the 4v4 balanced scenario. We varied \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda}_{d}\in \{\mathrm{0.3,0.5,0.8}\}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda}_{h}\in \{\mathrm{0.1,0.3,0.5}\}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mid {r}_{pen}/{r}_{cap}\mid \in \{\mathrm{2.0,2.5,3.0,5.0}\}$$\end{document}, evaluating each configuration over three seeds. The selected weights achieved the highest interception rate while keeping oscillation rate below 10%. We note that the global reward magnitudes (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{cap}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{pen}$$\end{document}) dominate shaping signals by design: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mid {r}_{pen}\mid =2.5\times {r}_{cap}$$\end{document} reflects the asymmetric operational cost of a defensive breach versus a successful interception. Moderate sensitivity was observed for the penalty-to-reward ratio: configurations with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mid {r}_{pen}/{r}_{cap}\mid =2.0$$\end{document} produced 2–3 point lower interception rates due to insufficient breach deterrence, while ratios above 3.0 yielded comparable results. Training procedure Training follows a two-stage strategy, summarized in Table 3. In the first stage, the high-level DT-GAT parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta}_{h}$$\end{document} are frozen, and allocation is replaced by uniform random assignment. This stage focuses on pretraining the low-level policy parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta}_{l}$$\end{document} and Critic parameters ϕ, so the maneuver network acquires basic target-tracking and interception capabilities. Establishing a reliable maneuvering baseline at this stage prevents low-level policy degradation that would otherwise arise from unstable assignments at the onset of joint training. Two independent Adam optimizers handle \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta}_{l}$$\end{document} and ϕ respectively.Table 3Hyperparameter settings including τ annealing schedule.ParameterStage 1Stage 2Training steps\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times {10}^{6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times {10}^{6}$$\end{document}High-level LR \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\eta}}}_{{\boldsymbol{h}}}$$\end{document}— (frozen)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\times {10}^{-5}$$\end{document}Low-level LR \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\eta}}}_{{\boldsymbol{l}}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times {10}^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times {10}^{-4}$$\end{document}Critic LR \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\eta}}}_{{\boldsymbol{\phi}}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times {10}^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times {10}^{-4}$$\end{document}Temperature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\tau}}$$\end{document}—5.0→ 0.5(linear)Allocation period \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{K}}$$\end{document}—20 stepsDiscount \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\gamma}}$$\end{document}0.990.99GAE \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\lambda}}$$\end{document}0.950.95Attention heads \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{H}}$$\end{document}—4Embedding dim \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{d}}$$\end{document}/ head dim \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{d}}}^{\mathbf{^{\prime}}}$$\end{document}64 / —64 / 16Rollout buffer\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\times {10}^{4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times {10}^{5}$$\end{document}Parallel envs33Batch size20484096 Table 3Hyperparameter settings including τ annealing schedule.ParameterStage 1Stage 2Training steps\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times {10}^{6}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times {10}^{6}$$\end{document}High-level LR \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\eta}}}_{{\boldsymbol{h}}}$$\end{document}— (frozen)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\times {10}^{-5}$$\end{document}Low-level LR \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\eta}}}_{{\boldsymbol{l}}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times {10}^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times {10}^{-4}$$\end{document}Critic LR \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{\eta}}}_{{\boldsymbol{\phi}}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times {10}^{-4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3\times {10}^{-4}$$\end{document}Temperature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\tau}}$$\end{document}—5.0→ 0.5(linear)Allocation period \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{K}}$$\end{document}—20 stepsDiscount \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\gamma}}$$\end{document}0.990.99GAE \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{\lambda}}$$\end{document}0.950.95Attention heads \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{H}}$$\end{document}—4Embedding dim \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\boldsymbol{d}}$$\end{document}/ head dim \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{d}}}^{\mathbf{^{\prime}}}$$\end{document}64 / —64 / 16Rollout buffer\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5\times {10}^{4}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times {10}^{5}$$\end{document}Parallel envs33Batch size20484096 Hyperparameter settings including τ annealing schedule. In the second stage, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta}_{h}$$\end{document} is unfrozen and both tiers train jointly. The Gumbel-Softmax temperature anneals linearly from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{0}=5.0$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\tau}_{\mathrm{m}\mathrm{i}\mathrm{n}}=0.5$$\end{document}. Three separate Adam optimizers update \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta}_{h}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta}_{l}$$\end{document}, and ϕ, with the high-level learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\eta}_{h}$$\end{document} set to half of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\eta}_{l}$$\end{document} to reduce the impact of allocation policy fluctuations on the already-established maneuvering capabilities. The complete two-stage training procedure is summarized in Algorithm 1. Algorithm 1: DT-GAT-MARL training procedure. Loss interaction and convergence. The two tiers are optimized by separate losses. The low-level Actor is updated via the clipped PPO surrogate:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{L}}^{l}({\theta}_{l})=-{\mathbb{E}}\left[min\left({\rho}_{t}{\widehat{A}}_{t},{\mathrm{clip}}({\rho}_{t},1-\epsilon ,1+\epsilon ){\widehat{A}}_{t}\right)\right]$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rho}_{t}={\pi}_{{\theta}_{l}}({a}_{t}\mid {o}_{t},{\widetilde{z}}_{i})/{\pi}_{{\theta}_{l}^{old}}({a}_{t}\mid {o}_{t},{\widetilde{z}}_{i})$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat{A}}_{t}$$\end{document} is the GAE advantage estimate. The Critic minimizes a value regression loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{L}}^{V}(\phi )={\mathbb{E}}[({V}_{\phi }(\mathcal{S})-{R}_{t}^{\lambda }{)}^{2}]$$\end{document}. The high-level DT-GAT parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta}_{h}$$\end{document} receive gradients through \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widetilde{z}}_{i}={\sum}_{j}{\pi}_{ij}{z}_{j}^{u}$$\end{document} via the chain rule \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\nabla}_{{\theta}_{h}}{\mathcal{L}}^{l}={\nabla}_{{\widetilde{z}}_{i}}{\mathcal{L}}^{l}\cdot {\nabla}_{{\pi}_{ij}}{\widetilde{z}}_{i}\cdot {\nabla}_{{\theta}_{h}}{\pi}_{ij}$$\end{document}, where the Gumbel-Softmax reparameterization ensures \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\nabla}_{{\theta}_{h}}{\pi}_{ij}$$\end{document} exists. The two-stage training strategy is designed to mitigate instability arising from this coupling. In Stage 1, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta}_{h}$$\end{document} is frozen and the low-level policy converges under random allocation, establishing a stable maneuvering baseline. In Stage 2, the halved high-level learning rate (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\eta}_{h}=0.5{\eta}_{l}$$\end{document}) limits the magnitude of allocation shifts per update, preventing catastrophic interference with the already-learned maneuvering policy. Fig. 3(b) shows that the training curve exhibits a transient performance dip during mid-Stage 2, corresponding to the temperature regime where allocation transitions from exploratory to near-deterministic; convergence resumes once τ drops sufficiently. Empirically, the joint training consistently converges across all five random seeds, with final performance variance within 3.5 percentage points. Experiments and analysis Setup and baselines The simulation environment operates in continuous 2D space with physical parameters from Table 1. Kinematic integration uses fourth-order Runge–Kutta (RK4) at Δ t=0.1 s, with episodes capped at 400 steps. All experiments ran on servers equipped with NVIDIA A100 GPUs. Table 4 details the scenario configurations.Table 4Simulation parameters and scenario configurations.Scenario\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{N}}}_{{\boldsymbol{d}}}\times {{\boldsymbol{N}}}_{{\boldsymbol{u}}}$$\end{document}Area (m)Initial deploymentDescriptionBalanced-S4 × 42000 × 2000Defenders bottom-random, intruders top-randomBaselineBalanced-M8 × 83000 × 3000SameMedium scaleBalanced-L12 × 124000 × 4000SameScalabilityDisadv.-Light4 × 62000 × 2000SameNumerical disadvantageDisadv.-Heavy6 × 103000 × 3000SameSevere disadvantageDynamic intrusion4 × 4 + (1 + 3)2000 × 2000Initial 4v4; 1 intruder at t=80, 3 at t=120Dynamic topology Table 4Simulation parameters and scenario configurations.Scenario\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{N}}}_{{\boldsymbol{d}}}\times {{\boldsymbol{N}}}_{{\boldsymbol{u}}}$$\end{document}Area (m)Initial deploymentDescriptionBalanced-S4 × 42000 × 2000Defenders bottom-random, intruders top-randomBaselineBalanced-M8 × 83000 × 3000SameMedium scaleBalanced-L12 × 124000 × 4000SameScalabilityDisadv.-Light4 × 62000 × 2000SameNumerical disadvantageDisadv.-Heavy6 × 103000 × 3000SameSevere disadvantageDynamic intrusion4 × 4 + (1 + 3)2000 × 2000Initial 4v4; 1 intruder at t=80, 3 at t=120Dynamic topology Simulation parameters and scenario configurations. All DRL baselines use the same low-level MAPPO Actor-Critic architecture (two-layer MLP with 128 hidden units and GRU with 64-dimensional hidden state) and identical observation/action spaces. GraphSAGE + MAPPO replaces DT-GAT with a two-layer GraphSAGE encoder using mean aggregation, operating on the same bipartite graph \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{G}}_{t}$$\end{document} with identical edge features but without the attention mechanism or dynamic masking. DGN uses a two-layer graph convolutional network with the same embedding dimension (d = 64). For methods without an explicit hierarchical structure (QMIX, MAPPO, DGN), allocation is derived from the policy network’s output at every timestep rather than at K-step intervals. Network architecture details. The low-level Actor consists of a two-layer MLP (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}_{o}\to 128\to 128$$\end{document}) followed by a GRU with 64-dimensional hidden state, and a final linear layer mapping to the action space \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(\Delta {v}_{i},{\omega}_{i}\right)$$\end{document}. The Critic uses a two-layer MLP (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}_{s}\to 256\to 256$$\end{document}) with no recurrence. The DT-GAT encoder has a single attention layer with H=4 heads, input dimension d=64, per-head dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}^{^{\prime}}=16$$\end{document}, and edge feature dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}_{f}=3$$\end{document}. The bilinear allocation head uses \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{b}\in {\mathbb{R}}^{64\times 64}$$\end{document}. All networks use ReLU activations except ELU in the GAT aggregation (Eq. 6). Training uses Adam with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\beta}_{1}=0.9$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\beta}_{2}=0.999$$\end{document}, and gradient clipping at max norm 10. Baselines include the Hungarian algorithm (optimal static matching), a rolling-horizon Hungarian variant (RH-Hungarian) that re-solves optimal matching every K=20 steps to match DT-GAT’s allocation frequency, greedy allocation (nearest-first), QMIX25, MAPPO24 (pure policy gradient without graph networks), GraphSAGE + MAPPO20 (graph aggregation but no attention or dynamic masking), and DGN26 (graph convolutional RL). We note that the Hungarian algorithm and greedy allocation serve as static optimization references rather than competitive dynamic baselines. To contextualize DT-GAT’s dynamic capability, we also report a rolling-horizon Hungarian (RH-Hungarian) variant that re-solves the optimal matching every K= 20 steps, matching DT-GAT’s allocation frequency. The Hungarian and RH-Hungarian baselines serve as optimization-based references: the former captures the best achievable one-shot static assignment, while the latter tests whether periodic re-optimization alone—without learned representations—can approach learning-based dynamic allocation. For QMIX, which natively targets discrete action spaces, we uniformly discretized each of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${v}_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\omega}_{i}$$\end{document} into 5 levels, yielding 25 individual actions per agent. All DRL baselines were trained in the same environment under identical budgets (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4\times {10}^{6}$$\end{document} total steps), with 5 random seeds per configuration and 50 evaluation episodes per seed (250 runs total). For DT-GAT-MARL, the first \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1\times {10}^{6}$$\end{document} steps pretrain only the low-level policy under random allocation (Stage 1), whereas baselines learn continuously throughout; the total gradient update count is therefore comparable, though the learning signals differ in the early phase. Metrics include interception rate (fraction of intruders successfully intercepted), breach rate, mean interception time, effective reallocation rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R}_{re}={N}_{cap}^{re}/{N}_{trig}$$\end{document}, and allocation oscillation rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${O}_{rate}={N}_{switch}^{void}/{N}_{trig}$$\end{document}. Here \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{trig}$$\end{document} counts total event-triggered reallocations, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{cap}^{re}$$\end{document} counts successful interceptions following reallocation, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{switch}^{void}$$\end{document} counts periodic triggers where the new assignment actually increased target distance. In the main comparison (Table 5), the latter two metrics are reported only for the dynamic intrusion scenario. Because each intruder is ultimately either intercepted or breached, the breach rate equals 100% minus the interception rate and is therefore omitted from Table 5 for brevity; in the ablation study (Table 6), oscillation rate also quantifies each module’s contribution to allocation stability.Table 5Performance comparison across all scenarios (mean ± std, 250 runs).Interception rate (%)Method4v48v812v124v66v10DynamicHungarian87.2±3.170.4±4.858.6±5.950.1±6.233.8±7.049.3±7.5RH-Hungarian88.5±3.073.6±4.561.2±5.753.4±5.937.2±6.855.8±7.0Greedy79.6±4.563.2±5.651.4±6.841.7±7.327.2±7.843.8±8.2QMIX83.4±3.869.8±5.157.2±6.447.3±6.534.1±7.451.6±7.1MAPPO89.1±3.277.6±4.267.3±5.556.8±5.842.5±6.957.2±6.8GraphSAGE91.3±2.881.4±3.972.1±5.161.7±5.447.3±6.561.4±7.2DGN90.5±3.079.8±4.170.5±5.359.2±5.745.6±6.763.8±6.5DT-GAT93.6±2.486.3±3.578.4±4.768.5±5.054.7±6.174.1±5.3Mean interception time (s)Method4v48v812v124v66v10DynamicHungarian12.3±1.818.7±2.926.4±3.815.1±2.422.8±3.520.6±3.2RH-Hungarian12.1±1.717.8±2.725.1±3.614.6±2.321.5±3.319.2±3.0Greedy15.4±2.422.1±3.430.5±4.518.8±3.127.3±4.125.0±3.9QMIX14.1±2.220.3±3.028.1±3.917.2±2.725.0±3.722.8±3.4MAPPO13.5±1.918.2±2.624.5±3.415.6±2.421.2±3.219.5±2.8GraphSAGE13.0±1.717.4±2.523.2±3.314.5±2.220.0±2.918.4±2.7DGN13.3±1.817.9±2.623.7±3.414.9±2.320.6±3.118.9±2.8DT-GAT12.8±1.516.2±2.321.7±3.213.4±2.118.9±3.017.3±2.6Effective reallocation rate\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{R}}}_{re}$$\end{document}(%) — dynamic intrusion onlyMethodDynamicQMIX52.3±7.8MAPPO58.4±7.1RH-Hungarian63.5±6.4GraphSAGE71.2±5.8DGN68.5±6.3DT-GAT87.3±4.2Allocation oscillation rate\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{O}}}_{rate}$$\end{document}(%) — dynamic intrusion onlyMethodDynamicQMIX25.1±4.5MAPPO22.3±4.0RH-Hungarian5.2±2.0GraphSAGE15.7±3.4DGN17.4±3.7DT-GAT9.6±2.6 Table 5Performance comparison across all scenarios (mean ± std, 250 runs).Interception rate (%)Method4v48v812v124v66v10DynamicHungarian87.2±3.170.4±4.858.6±5.950.1±6.233.8±7.049.3±7.5RH-Hungarian88.5±3.073.6±4.561.2±5.753.4±5.937.2±6.855.8±7.0Greedy79.6±4.563.2±5.651.4±6.841.7±7.327.2±7.843.8±8.2QMIX83.4±3.869.8±5.157.2±6.447.3±6.534.1±7.451.6±7.1MAPPO89.1±3.277.6±4.267.3±5.556.8±5.842.5±6.957.2±6.8GraphSAGE91.3±2.881.4±3.972.1±5.161.7±5.447.3±6.561.4±7.2DGN90.5±3.079.8±4.170.5±5.359.2±5.745.6±6.763.8±6.5DT-GAT93.6±2.486.3±3.578.4±4.768.5±5.054.7±6.174.1±5.3Mean interception time (s)Method4v48v812v124v66v10DynamicHungarian12.3±1.818.7±2.926.4±3.815.1±2.422.8±3.520.6±3.2RH-Hungarian12.1±1.717.8±2.725.1±3.614.6±2.321.5±3.319.2±3.0Greedy15.4±2.422.1±3.430.5±4.518.8±3.127.3±4.125.0±3.9QMIX14.1±2.220.3±3.028.1±3.917.2±2.725.0±3.722.8±3.4MAPPO13.5±1.918.2±2.624.5±3.415.6±2.421.2±3.219.5±2.8GraphSAGE13.0±1.717.4±2.523.2±3.314.5±2.220.0±2.918.4±2.7DGN13.3±1.817.9±2.623.7±3.414.9±2.320.6±3.118.9±2.8DT-GAT12.8±1.516.2±2.321.7±3.213.4±2.118.9±3.017.3±2.6Effective reallocation rate\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{R}}}_{re}$$\end{document}(%) — dynamic intrusion onlyMethodDynamicQMIX52.3±7.8MAPPO58.4±7.1RH-Hungarian63.5±6.4GraphSAGE71.2±5.8DGN68.5±6.3DT-GAT87.3±4.2Allocation oscillation rate\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\boldsymbol{O}}}_{rate}$$\end{document}(%) — dynamic intrusion onlyMethodDynamicQMIX25.1±4.5MAPPO22.3±4.0RH-Hungarian5.2±2.0GraphSAGE15.7±3.4DGN17.4±3.7DT-GAT9.6±2.6 Performance comparison across all scenarios (mean ± std, 250 runs). Comparative results Table 5 presents comprehensive performance comparisons across six scenario configurations. In the 4v4 balanced scenario, all DRL methods achieve relatively high interception rates, with DT-GAT (93.6%) outperforming MAPPO (89.1%) by 4.5 percentage points. Across all six scenarios, DT-GAT’s interception rate improvements over the strongest baseline in each scenario are statistically significant by Welch’s t-test. Table 5 reports mean ± std over 250 runs; specific p-values are: 4v4 p=0.032, 8v8 p=0.003, 12v12 p=0.001, 4v6 p=0.004, 6v10 p=0.006, Dynamic p<0.001. The 95% confidence intervals for DT-GAT’s advantage over the next-best method range from [0.8, 4.1] percentage points in the 4v4 scenario to [7.4, 13.2] points in the dynamic intrusion scenario. It is worth noting that the Hungarian algorithm edges out DT-GAT on mean interception time in this setting (12.3 s vs. 12.8 s)—optimal static matching produces near-globally-optimal initial assignments when the problem is small and symmetric. As the scale grows to 8v8 and 12v12, however, the Hungarian algorithm’s lack of dynamic adaptability becomes apparent: its success rate drops from 87.2% to 58.6%, whereas DT-GAT declines only from 93.6% to 78.4%, exhibiting notably better scalability. The RH-Hungarian baseline, which re-optimizes matching every K steps, consistently outperforms the static Hungarian algorithm across all scenarios (e.g., 73.6% vs. 70.4% at 8v8), confirming that periodic re-optimization is strictly beneficial. However, even with identical allocation frequency and event triggers, RH-Hungarian falls well short of DT-GAT (73.6% vs. 86.3% at 8v8)—a gap of 12.7 points attributable entirely to the learned allocation policy. This comparison isolates the value of end-to-end learned attention-based assignment: whereas RH-Hungarian minimizes instantaneous distances, DT-GAT’s edge-feature biases integrate distance, bearing, and estimated intercept time to produce allocations that are geometrically informed rather than myopically optimal. That said, the 78.4% rate at 12v12 indicates that large-scale cooperative interception remains challenging for current methods. Among graph-based approaches, DT-GAT surpasses GraphSAGE + MAPPO by 4.9 points in the 8v8 scenario (86.3% vs. 81.4%), a gap attributable primarily to the spatial awareness conferred by edge-feature biases and the adaptive aggregation from the attention mechanism. DGN performs comparably to GraphSAGE, though its uniform graph convolution aggregation limits differentiated focus when intruder distributions are uneven. Disadvantaged scenarios amplify these differences further: under the severe 6v10 imbalance, DT-GAT (54.7%) exceeds MAPPO (42.5%) by 12.2 points, as the release-and-reassign mechanism enables interceptors to pivot promptly toward fresh threats after completing an interception. Even so, more than 40% of intruders breach the defense in the 6v10 setting, underscoring the inherent difficulty of numerically disadvantaged scenarios. Dynamic re-allocation analysis The dynamic intrusion scenario serves as the primary validation for DT-GAT, as depicted in Fig. 2. Starting from a 4v4 engagement, four new intruders enter in two waves. DT-GAT achieves 74.1% interception in this scenario, outperforming the next-best baseline DGN (63.8%) by 10.3 points—the largest margin observed across all configurations. The effective reallocation rate more directly reflects dynamic adaptability: DT-GAT reaches 87.3%, compared with 71.2% for GraphSAGE and 58.4% for MAPPO. On oscillation rate, DT-GAT records 9.6%, well below GraphSAGE’s 15.7% and MAPPO’s 22.3%, confirming that the switching cost \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}_{switch}$$\end{document} and the dual-frequency mechanism effectively suppress unproductive oscillations. Notably, RH-Hungarian achieves an even lower oscillation rate of 5.2% owing to its deterministic matching, yet its interception rate (55.8%) lags DT-GAT by 18.3 points. This contrast highlights that minimal oscillation is not inherently desirable—some allocation switches reflect valuable proactive reallocation toward emerging threats rather than unproductive churn. DT-GAT strikes a more effective balance between allocation stability and adaptive responsiveness.Fig. 2Performance and attention dynamics in the dynamic intrusion scenario. Performance and attention dynamics in the dynamic intrusion scenario. Figure 2(b) traces the attention weight evolution of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{D}}_{1}$$\end{document} during a representative engagement. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{D}}_{1}$$\end{document} initially concentrates attention on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{U}}_{1}$$\end{document} and, upon successful interception at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{t}$$\end{document}=60, redirects toward \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{U}}_{2}$$\end{document}. When a new target \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{U}}_{5}$$\end{document} appears at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{t}$$\end{document}=80, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{D}}_{1}$$\end{document}’s attention to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{U}}_{5}$$\end{document} rises to roughly 0.3, yet because \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{U}}_{2}$$\end{document} remains closer, the allocation decision keeps \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{D}}_{1}$$\end{document} on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{U}}_{2}$$\end{document}—a behavior consistent with the edge-feature bias integrating distance and urgency. In a few episodes, brief attention fluctuations among multiple new targets were observed, suggesting that allocation stability under dense event triggering still has room for improvement. Ablation studies Table 6 reports results for seven ablation configurations in the 8v8 scenario, isolating each module’s individual contribution.Table 6Ablation experiment results (8v8 scenario, 250 runs).ConfigDescriptionSuccess (%)Breach (%)Oscillation (%)AFull DT-GAT-MARL86.3 ± 3.510.2 ± 3.48.2 ± 2.1BFixed-rule allocation (Hungarian replaces DT-GAT)74.8 ± 4.620.1 ± 4.3—CNo masking (rebuild graph on topology change)83.7 ± 3.912.8 ± 3.69.1 ± 2.4DNo edge-feature bias (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\upbeta}_{\mathrm{k}}=0$$\end{document})80.1 ± 4.215.6 ± 3.911.5 ± 2.8ENo Gumbel-Softmax (plain softmax)82.4 ± 4.013.3 ± 3.713.8 ± 3.2FNo hierarchy (single-tier end-to-end)78.6 ± 4.517.2 ± 4.116.4 ± 3.5GAdded manual cooperation reward85.1 ± 3.711.4 ± 3.59.5 ± 2.5Note: Success rate counts intruders intercepted within the capture radius; breach rate counts intruders that crossed the defense-zone boundary. Intruders still in flight at the episode time limit account for the residual fraction. Table 6Ablation experiment results (8v8 scenario, 250 runs).ConfigDescriptionSuccess (%)Breach (%)Oscillation (%)AFull DT-GAT-MARL86.3 ± 3.510.2 ± 3.48.2 ± 2.1BFixed-rule allocation (Hungarian replaces DT-GAT)74.8 ± 4.620.1 ± 4.3—CNo masking (rebuild graph on topology change)83.7 ± 3.912.8 ± 3.69.1 ± 2.4DNo edge-feature bias (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\upbeta}_{\mathrm{k}}=0$$\end{document})80.1 ± 4.215.6 ± 3.911.5 ± 2.8ENo Gumbel-Softmax (plain softmax)82.4 ± 4.013.3 ± 3.713.8 ± 3.2FNo hierarchy (single-tier end-to-end)78.6 ± 4.517.2 ± 4.116.4 ± 3.5GAdded manual cooperation reward85.1 ± 3.711.4 ± 3.59.5 ± 2.5Note: Success rate counts intruders intercepted within the capture radius; breach rate counts intruders that crossed the defense-zone boundary. Intruders still in flight at the episode time limit account for the residual fraction. Ablation experiment results (8v8 scenario, 250 runs). Note: Success rate counts intruders intercepted within the capture radius; breach rate counts intruders that crossed the defense-zone boundary. Intruders still in flight at the episode time limit account for the residual fraction. Replacing DT-GAT with the Hungarian algorithm (Config B) drops the success rate by 11.5 points (74.8% vs. 86.3%), confirming the advantage of end-to-end learned allocation over fixed rules. Removing masking (Config C) causes only a 2.6-point decline in the 8v8 balanced scenario (83.7%), which is expected since this scenario involves no node dynamics—the masking mechanism’s value manifests primarily in dynamic scenarios, where the degradation is far more pronounced. Config D, which zeroes out the edge-feature bias, yields the steepest single-module drop at 6.2 points (80.1%), identifying this component as the most critical design element in DT-GAT; without it, attention weights degrade to topology-agnostic aggregation. Removing Gumbel-Softmax (Config E) reduces the success rate to 82.4% while the oscillation rate climbs to 13.8%, highlighting the non-trivial roles of Gumbel noise and temperature annealing in exploring diverse assignments and stabilizing convergence. The result of dismantling the hierarchical architecture (Config F) merits closer examination. The success rate falls to 78.6% and oscillation rate spikes to 16.4%, because updating allocation and maneuvering at the same timescale allows per-step assignment fluctuations to directly disrupt maneuver continuity. Config G, which adds a handcrafted cooperative reward, achieves 85.1%—slightly below the full method’s 86.3%, a difference not statistically significant (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{p}>0.05$$\end{document})—yet its oscillation rate edges upward. From this we can infer that supplementary cooperative signals did not yield the expected benefit and may even introduce reward conflicts, vindicating the design principle of relying primarily on task-level signals. Sensitivity and interpretability Figure 3(a) shows the influence of the allocation period \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{K}$$\end{document} on performance. Peak success occurs at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{K}$$\end{document}=20; overly frequent updates (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{K}$$\end{document}=5) drive oscillation to 28.4% and lower success to 81.2%, while overly infrequent updates (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{K}$$\end{document}=50) yield sluggish responses and 76.5% success. Performance at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{K}$$\end{document}=10 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{K}$$\end{document}=30 remains within a reasonable band, indicating moderate robustness to this parameter. Figure 3(b) compares three temperature strategies: annealing (τ: 5.0 → 0.5) achieves the highest success at 86.3%, while fixed τ=1.0 and τ=0.5 converge to 82.1% and 79.8%, respectively. A transient dip during mid-training under the annealing schedule corresponds to the transition from exploratory to near-deterministic allocation.Fig. 3Sensitivity analysis and computational efficiency. Sensitivity analysis and computational efficiency. Figure 3(c) plots inference time against the total number of agents. At 12v12 (24 agents), inference takes approximately 6.8 ms, rising to about 10.2 ms at 40 agents. Given that perception, communication, and execution share the control period (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta \mathrm{t}=0.1$$\end{document} s), decision-making inference typically needs to remain within 10% of the timestep27 (i.e., 10 ms, shown as a dashed line in Fig. 3(c)) to leave adequate communication and execution margins; current inference times already approach this bound. GraphSAGE, which forgoes attention weight computation, runs faster at the same scale (~ 7.5 ms at 40 agents), revealing the computational cost of DT-GAT’s attention mechanism at large scales. Figure 4 presents attention weight matrices at three key moments during a representative episode28: initially near-diagonal, reflecting distance preferences; after new targets appear, attention redistributes across multiple targets as distance and urgency compete; once released, attention swiftly concentrates on the new target, indicating decisive reallocation.Fig. 4Attention weight matrices of a representative episode in the dynamic intrusion scenario. Attention weight matrices of a representative episode in the dynamic intrusion scenario. Computational complexity The dominant cost of DT-GAT lies in the bipartite attention computation. For \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{N}}_{\mathrm{d}}$$\end{document} defenders and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{N}}_{\mathrm{u}}$$\end{document} active intruders with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{H}$$\end{document} attention heads and embedding dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{d}$$\end{document}, the per-layer complexity of edge-feature-enhanced attention is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{O}(\mathrm{H}\cdot {\mathrm{N}}_{\mathrm{d}}\cdot {\mathrm{N}}_{\mathrm{u}}\cdot \mathrm{d})$$\end{document}, arising from computing all pairwise attention logits (Eq. 3) and the subsequent weighted aggregation (Eq. 6). Edge feature extraction adds \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{O}({\mathrm{N}}_{\mathrm{d}}\cdot {\mathrm{N}}_{\mathrm{u}}\cdot {\mathrm{d}}_{\mathrm{f}})$$\end{document}, which is subsumed by the attention term since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{d}}_{\mathrm{f}}\ll \mathrm{d}$$\end{document}. The Gumbel-Softmax allocation head contributes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{O}({\mathrm{N}}_{\mathrm{d}}\cdot {\mathrm{N}}_{\mathrm{u}}\cdot \mathrm{d})$$\end{document} for the bilinear scoring. On the low-level side, the MAPPO Actor-Critic forward pass is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{O}({\mathrm{N}}_{\mathrm{d}}\cdot {\mathrm{d}}_{net}^{2})$$\end{document} per timestep, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{d}}_{net}$$\end{document} is the hidden dimension. Since the high-level module operates every \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{K}$$\end{document} steps, the amortized per-step complexity is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{O}({\mathrm{N}}_{\mathrm{d}}\cdot {\mathrm{N}}_{\mathrm{u}}\cdot \mathrm{H}\cdot \mathrm{d}/\mathrm{K}+{\mathrm{N}}_{\mathrm{d}}\cdot {\mathrm{d}}_{net}^{2})$$\end{document}. The quadratic dependence on agent count (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{N}}_{\mathrm{d}}\cdot {\mathrm{N}}_{\mathrm{u}}$$\end{document}) is the primary scalability bottleneck, consistent with the inference time growth observed in Fig. 3(c). Conclusion This paper has presented DT-GAT-MARL, a dual-frequency hierarchical decision framework for dynamic interception task allocation in counter-drone defense. The DT-GAT module combines masking, edge-feature biases, and a Gumbel-Softmax allocation head to achieve differentiable end-to-end assignment under dynamic topologies. Experimental results pinpoint the edge-feature bias as the single most impactful component (6.2-point ablation drop), while masking ensures computational continuity during topology changes in dynamic scenarios, and Gumbel-Softmax temperature annealing balances exploration against convergence stability. The dual-frequency mechanism effectively suppresses allocation oscillation by decoupling assignment and maneuvering timescales (oscillation rate rose from 8.2% to 16.4% when hierarchy was removed), and DT-GAT outperforms the best baseline by 10.3 points in the dynamic intrusion scenario. Several limitations warrant acknowledgment. The kinematic model is a planar three-variable simplification that omits three-dimensional maneuvers and aerodynamic constraints. Intruders follow constant-velocity straight-line trajectories representing a penetration-focused strategy; in real engagements, adversarial drones are likely to exhibit evasive or cooperative behaviors29, which would substantially raise both interception difficulty and the demands on the allocation module. Similarly, the planar kinematic model omits altitude dynamics and aerodynamic constraints present in three-dimensional operational environments. DT-GAT’s inference time grows relatively fast with agent count (~ 10.2 ms at 40 agents), and scalability is bounded by the quadratic complexity of attention computation30. Furthermore, the current dynamic topology handling relies on masking over a fixed-capacity graph rather than true dynamic graph learning methods such as temporal graph networks. While masking is computationally efficient and avoids structural reconstruction, it does not learn temporal transition patterns between successive graph states. Incorporating dynamic graph neural networks that explicitly model topology evolution over time could potentially improve anticipatory reallocation. The 78.4% success rate at 12v12 further points to considerable room for improvement in large-scale coordination. Future work will pursue several directions: introducing sparse attention mechanisms to reduce computational overhead at scale; investigating decentralized allocation strategies based on local subgraphs under communication constraints31; modeling intruders as intelligent evasive agents within an adversarial game-theoretic training framework; and exploring sim-to-real transfer methods32 for deploying trained policies on physical drone platforms.