Multi-agent reinforcement learning driven resource game optimization for network slicing in MEC-enabled HetNets Authors: Kai Mao Published: 2025 ABSTRACT AbstractThe convergence of Mobile Edge Computing (MEC) and network slicing technologies is critical to meet the diverse quality of service (QoS) requirements of 5G/6G.However,multi-tenancy (eMBB/uRLLC/mMTC) competition for resources in dynamic environments challenges traditional centralised allocation methods.In this paper, we propose a cooperative optimization framework for edge network slicing resources based on the fusion of multi-intelligence reinforcement learning (MARL) and evolutionary game theory (MARL-EGT).The framework models each slice tenant as an intelligent body with autonomous decision-making capability, which explores the optimal resource requesting strategy through interactive learning; meanwhile, evolutionary game dynamics is introduced to model the imitation, learning and evolution process of the slice population strategy, which guides the system to converge to an efficient evolutionary stable equilibrium (ESS).In order to cope with the problem of too large environment state space and intelligence coordination, a hierarchical attention mechanism and a credit-based contribution evaluation algorithm are innovatively designed to significantly improve the learning efficiency and convergence speed. In simulation experiments, under the MEC scenario constructed based on real data, the MARL-EGT scheme significantly outperforms benchmark methods such as federated reinforcement learning (FRL) and non-cooperative gaming (NCG) in terms of key metrics, such as total system utility, slicing SLA satisfaction rate, and resource utilization, and demonstrates superior dynamic environment adaptability, which provides large-scale, intelligent edge network slicing resource management new ideas.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-33190-5.Keywords: Mobile edge computing (MEC), Network slicing, Resource allocation, Multi-intelligence reinforcement learning (MARL), Evolutionary game theory (EGT), Quality of service (QoS), Intelligent optimizationSubject terms: Engineering, Mathematics and computing Introduction Multi-agent reinforcement learning (MARL), as a cross research field of game theory, cybernetics and multi-intelligent learning, is a cutting-edge direction in the research of multi-agent systems (MASs), which empowers intelligences to complete diversified tasks in dynamic and multi-dimensional complex environments through interaction and decision-making. MASs are at the forefront of multi-agent systems (MASs) research, empowering intelligences to perform diverse tasks in dynamic, multi-dimensional and complex environments through interaction and decision-making. The emergence of network slicing enables mobile networks to cope with different demands without the need to set up a physical network, and at the same time further strengthens the resource management capability in mobile networks, reduces the cost and improves the network performance. However, since the network environment is not static, network resources need to be allocated adaptively to cope with different situations; at the same time, there are various terminal devices with different limitations and requirements; furthermore, multiple slices interact with each other, resulting in the need for further allocation and adjustment of network resources. Network slicing technology becomes a key enabling technology to meet diverse quality of service (QoS) requirements by constructing multiple customised virtual networks with specific performance metrics in logical isolation on a set of shared physical network infrastructure. Overview of related work Existing research on resource optimization for MEC network slicing revolves around two main lines: centralized optimization and distributed coordination. Centralized approaches usually use mathematical tools such as convex optimization, Liapunov optimization, and column generation to model the resource allocation problem. Wenbin Yao modeled slicing resource allocation as a stochastic optimization problem, using Liapunov drift plus penalty function to minimize the long-term average cost and delay1.Qingzheng Tan proposed an adaptive deflation algorithm for core network elements in response to network resource requests over time and the need to further optimize network slicing performance. Using deep reinforcement learning to solve the optimal policy, and modeling and analysis of the core network user-functional network element processing system, the delay is reduced by 20% at the cost of 10% energy consumption. At the same time, imitation learning is added to the deep reinforcement learning algorithm to avoid the intelligence from falling into the wrong policy by using the traditional algorithm as the expert policy assistance for convergence. Meanwhile, for the resource allocation problem of multi-slice coexistence faced in 5G mobile networks, a resource allocation algorithm for multi-slice coexistence under the constraint of base station energy consumption is proposed from the perspective of wireless resources to optimize the slice performance. Taking the case of coexistence of throughput-sensitive slices and delay-sensitive slices as an example, the objective is to maximize the throughput of the throughput-sensitive slices under the condition of satisfying the performance of the delay-sensitive slices. A bandwidth allocation algorithm is proposed, and then it is combined with the semi-positive constraint algorithm to solve the optimization problem. By comparing with the traditional bandwidth allocation algorithm, it is proved that the proposed bandwidth allocation algorithm can improve the throughput by 7% compared with the average bandwidth allocation; and the overall algorithm is verified to increase the throughput by more than 6% compared with the genetic algorithm2. Based on the joint resource allocation problem of end-to-end network slicing with cloud-edge collaboration, a resource management and computation offloading scheme is proposed for the power consumption cost of slicing system. Firstly, a two-layer architecture model combining network slicing and mobile edge computing is established, which is mainly divided into slicing instance layer and infrastructure layer. Then, the optimization problem of minimizing the power cost is established by considering the communication and computation resources on the premise of guaranteeing the delay and rate of all users in the slices. Next, the complex problem is decomposed into two subproblems, and resource management and computation offloading algorithms based on hybrid penalty function and convex approximation are designed. Finally, simulation work is carried out, and the results verify the feasibility and effectiveness of the proposed scheme, which greatly reduces the system power cost, and validate the relationship between the number of users, the amount of MEC computational resources, the maximum base station transmit power, and the probability of packet error and the system power cost, respectively3. Multi-intelligence Reinforcement Learning (MRL) is developing in the direction of open application objects, embodied application problems, and complex application scenarios, and is gradually becoming the most effective tool for solving real-world game decision-making problems.The existence of Nash equilibrium is proved by modeling resource competition among slicing tenants using non-cooperative games. Ejaz A M, Wu G, Ahmed A designed an end-to-end network slicing approach to predict common and urgent user requests by T-distribution. We formulate the problem as a Multi-Agent Markov Decision Process (MDP) and introduce a Multi-Agent Soft Behavioural Critique (MAgSAC) algorithm. The algorithm balances the allocation process by intelligently activating and deactivating Virtual Network Function (VNF) instances to prevent wastage of scarce resources. Our approach aims to optimize the overall utility, balancing the trade-offs between revenue, energy cost and latency. We evaluate our approach MAgSAC through simulations and compare it to the following six benchmark scenarios: MAA3C, SACT, DDPG, S2Vec, Random and Greedy. The results show that our MAgSAC approach optimizes the utility by 30%, reduces the energy cost by 12.4% and shortens the execution time by 21.7% compared to the closest relevant multi-agent approach, MAA3C4. Existing solutions usually rely on static network resource partitioning or systematic a priori knowledge, which cannot adapt to the dynamic and changeable MEC environment, leading to problems such as excessive service delay and unreasonable resource provisioning. To address the above important challenges, Yu Z, Xu X, Zhou W proposed two resource allocation methods for 5G network slicing in MEC environment5.Jiayi L, Wenbin Y, Chen W considered a CDN slice system operating in discrete time intervals by adopting a periodic reconfiguration strategy. Within each time interval, the CDN-MEC-VNFs planning (CMVP) problem is formalized based on uncertainty programming to cope with dynamic and uncertain mobile network environments where user requests are treated as random variables with arbitrary patterns. Then, a data-driven and learning-based algorithm that integrates stochastic simulation (SS), deep neural network (DNN) and meta-heuristic algorithms is proposed for determining the provisioning of CDN slices based on historical user request information6.Foko L M S, Mthulisi V, Bertrand A B Combining SDN, NFV, SFC, and NS technologies toeffectively manage MEC server resources to ensure QoS requirements for AVNET. QoS parameters considered include latency, compute, storage, and bandwidth resources. Therefore, a mathematical resource management model for MEC servers is proposed and then, a new MEC architecture adapted to AVNET is proposed which uses the above mentioned techniques as well as the proposed mathematical model to manage bandwidth, compute and storage resources7. The explosive growth of mobile users (MUs) has led to a substantial increase in computing tasks being offloaded to MEC, M Chen, Y Zheng, J Yang, L Yang, W Zhang proposed the joint optimization problem of resource allocation, computation offloading, and resource pricing for multi-BS cooperative MEC in heterogeneous multicell networks8. And Krishnamoorthy S, Ramesh L, Subbarayalu R, et al. proposed model for capturing both the spatial and temporal features in network traffic patterns and allow accurate malicious pattern classification9. Additionally, in order to minimize the total energy consumption of the system and path loss of users on the premise. Long C, Jigang W, Jun Z, et al. developed an on-line offloading decision and computational resource management algorithm with joint consideration of collaborations between device–cloud, edge–edge and edge–cloud10, C Li, Y Gan, Y Zhang, Y Luo, et al. proposed a cooperative computation offloading strategy with on-demand deployment of multi-UAV in UAV-aided MEC11. System modeling and problem formalization Network architecture Consider a MEC system consisting of geographically distributed Edge Servers (ES) connected to the core network via backhaul links.The system supports $K$-class network slices (eMBB/uRLLC/mMTC), and each slice tenant (STₖ) provides services with specific SLA guarantees for its users.User equipment (UE) accesses the corresponding slices according to service types, forming an end-to-end virtualised network. Resource modeling Computational resources: each ES m possesses a limited computational capacity (CPU cycles/second) C_m. Communication resource: each ES m possesses a limited bandwidth resource (MHz) B_m. Multi-dimensional resource pool: the total system resources are represented as the vector R. Framework design Intelligent body design Figure 1 depicts an end-to-end network slicing framework for Multi-access Edge Computing (MEC). The core of the multi-intelligent body reinforcement learning framework lies in the observation-decision-reward closed-loop design of intelligent bodies. Each sliced intelligent body (eMBB/URLLC/mMTC) constructs an independent decision-making unit, and its implementation details are as follows: End-to-end network slicing framework for MEC enablement. The local observation space uses a tensor structure Queue state includes real-time task backlog (in MB), Poisson arrival rate (tasks/s), average task size (CPU cycles); Channel quality includes signal-to-noise ratio (dB), channel quality indicator (0–15), Doppler shift (Hz); Historical resource usage covers average CPU/bandwidth/memory utilisation over a sliding window (T = 10). Action space defined as continuous 3D vectors mapped to actual resource requests by linear transformation The action output layer uses the Tanh activation function to ensure that the value range[-1,1], and by affine transformations normalise to[0,1]. The reward function calculation process consists of three key steps Utility Calculation; Resource Waste Detection; and Final Reward. Reward computation is implemented in PyTorch with gradient backpropagation via automatic differentiation:. # Step 1: Calculate the base utility, if slice_type == “eMBB”. # Step 2: Resource pricing calculation. # Step 3: Waste detection. # Combined rewards. Distributed learning based on VDNs Value Decomposition Network (VDN) achieves synergy between distributed decision making and centralized training, and its architecture contains three types of neural networks: Strategy network (Actor)Input layer: 11-dimensional observation vector;Hidden layer: 2 layers of full connectivity (64 neurons, ReLU activation);Output layer: 3-dimensional action vectors (Tanh activation). Strategy network (Actor) Input layer: 11-dimensional observation vector; Hidden layer: 2 layers of full connectivity (64 neurons, ReLU activation); Output layer: 3-dimensional action vectors (Tanh activation). Value function network (Critic)Input layer: observation vector + action vector (14 dimensions);Hidden layer: 2 layers of full connectivity (128 neurons, ReLU activation);Output layer: scalar Q-value. Value function network (Critic) Input layer: observation vector + action vector (14 dimensions); Hidden layer: 2 layers of full connectivity (128 neurons, ReLU activation); Output layer: scalar Q-value. Mixed Network (Mixer).Realization value decomposition .Input: all intelligences Q-value splice vector ;Output: Joint action value ;The training algorithm uses a two-stage update mechanism:graph TB.A[Collection of empirical data] --> B[Updating the Critic Network of Intelligent Bodies];B --> C[VDN hybrid network meter Q_total];C --> D[Calculation of TD targets y = r + γQ’_total];D --> E[Minimizing TD errors L= (y-Q_total)²];E --> F[Strategy Gradient Update Actor Networks]. Mixed Network (Mixer). Realization value decomposition . Input: all intelligences Q-value splice vector ; Output: Joint action value ; The training algorithm uses a two-stage update mechanism: graph TB.A[Collection of empirical data] --> B[Updating the Critic Network of Intelligent Bodies];B --> C[VDN hybrid network meter Q_total];C --> D[Calculation of TD targets y = r + γQ’_total];D --> E[Minimizing TD errors L= (y-Q_total)²];E --> F[Strategy Gradient Update Actor Networks]. A[Collection of empirical data] --> B[Updating the Critic Network of Intelligent Bodies]; A[Collection of empirical data] --> B[Updating the Critic Network of Intelligent Bodies]; B --> C[VDN hybrid network meter Q_total]; B --> C[VDN hybrid network meter Q_total]; C --> D[Calculation of TD targets y = r + γQ’_total]; C --> D[Calculation of TD targets y = r + γQ’_total]; D --> E[Minimizing TD errors L= (y-Q_total)²]; D --> E[Minimizing TD errors L= (y-Q_total)²]; E --> F[Strategy Gradient Update Actor Networks]. E --> F[Strategy Gradient Update Actor Networks]. The specific settings for the hyperparameters are shown in Table 1: Table 1Hyper parameter configuration.ParameterPriceRoleDiscount factor γ0.95Balancing immediate/future rewardsLearning rate α5e-4Adam optimiser step sizeTarget update rate τ0.01Target network soft update factorExploration rate ε0.1→0.01Linear Decay Exploration NoiseBatch size32Number of empirical playback samplesBuffer size1e5Experience storage capacity Hyper parameter configuration. Proof of convergence of the game equilibrium To verify the theoretical reliability of the framework, we rigorously prove that the MARL strategy update converges to the Nash equilibrium through four steps, with supplementary mathematical deductions and condition verifications. Step 1: Establish equivalence of virtual games. Define the strategy update process of each slice agent (SIkik∈∈{1,2,3}) corresponding to eMBB/uRLLC/mMTC) as a discrete-time iterative system: where denotes the strategy distribution of agent i at iteration t; =argmaxa.i. ∈Ai Eπt−it [Ui (ai ,a−i)] is the best response of agent i to the strategies of other agents (πt−i​); I{·} is the indicator function (1 if the condition holds, 0 otherwise). The best response Equivalent to strategy gradient update: Where J(θi​) = Eπθi[Ui] is the expected utility of agent i, and Qi​(oi,a.i.) is the action-value function. When the learning rate satisfies , Strategy updates converge to a virtual game process. Convergence condition for strategy update: When the learning rate αt satisfies the Robbins-Monro conditions, where each iteration’s strategy is a best response to the previous iteration’s collective strategy of other agents. Step 2: Prove the existence of a Nash equilibrium. Tectonic resource gaming : Strategic Spatial Tightness , It is a closed-border subset of the European space; Utility function continuity with respect to Continuous (Price function Lipschitz continuous); Convexity of the fit: verification by Hessian matrices, Negative definite matrices guarantee concavity of the fit and Nash equilibrium exists according to Glicksberg’s theorem. Step 3: Convergence analysis. Strategy update generates a sequence of need certificates: Constructing compression mappings using Banach’s immovable point theorem and verifying the Lipschitz condition༚ Observed in the simulation (Calculated by Jacobian matrix), satisfy the convergence condition. Step 4: Experimental verification. Record the offset of the strategy from the equilibrium point during training: Table 2 documents the policy offsets and system Q-values across different training iterations, constituting the core data for convergence verification experiments. Table 2Number of training rounds.training roundδ_k (eMBB)δ_k (URLLC)δ_k (mMTC)Q_total00.850.920.78-12.310000.420.510.3718.750000.150.220.1168.510,0000.030.050.02102.1 Number of training rounds. When training rounds > 8000, and the standard deviation of the fluctuations < 0.01, showing convergence to equilibrium. The key performance test results of the framework are presented in Table 3. The end-to-end decision latency is merely 0.84 milliseconds, meeting the 5G TTI requirement. Compared with centralised RL and MADDPG, MARL-GTRO reduces communication overhead by 66.4% and 29.6% respectively, fully demonstrating the advantages of the VDN linear decomposition architecture. Table 3Key performance indicators.Decision latency analysisAssemblyDelay(ms)Accelerated Optimisation SolutionsObservation Coding0.28Operator FusionAction Generation0.41Layer Normalised CacheVDN Mixing0.15Parallel SummationTotals0.84< 1msCommunication overheadArchitectureSignalling overhead (Mbps)Comparison baselineCentralised RL45.6-MADDPG32.1-29.6 per centMARL-GTRO15.3-66.4 per cent Key performance indicators. The framework performance validation process deploys MARL-GTRO on the Intel Xeon Gold 6248R platform to test key performance indicators: The linear decomposition structure of VDN reduces the communication overhead by 66.4% compared to MADDPG, and the end-to-end decision delay of 0.84ms meets the 5G TTI requirements.The system utility is improved by 32% after training convergence, while the SLA default rate is reduced to less than 3%; Nash equilibrium convergence is rigorously proved theoretically and experimentally verified with offset δ_k < 0.05; the resource wastage rate is reduced to 5.1% from 22.7% in the baseline, verifying the effectiveness of the reward function. Experimentation and evaluation Simulation environment implementation The experiment is built based on Python 3.9, and the core dependencies include PyTorch 1.12 (neural network), OpenAI Gym 0.26 (environment interaction) and NS-3 3.36 (network simulation). The MEC server is configured with 32-core CPU/64G RAM/10G bandwidth. The core parameters for the three slicing services (URLLC, eMBB, mMTC) including task arrival rate, SLA requirements, and task characteristics are presented in Table 4. Table 4Parameters of the three types of slicing operations.Slicing typeMission arrival rateSLA RequirementsTask CharacteristicsURLLCΛ=-~Poisson(120/s)Latency < 10msPacket size: 32KB, CPU requirement: 500 cycles/pkteMBBλ ~ Poisson(40/s)Throughput > 1GbpsStream Duration: Exp(30s), Bandwidth Requirement: 50-200Mbps/streammMTCλ ~ Poisson(300/s)Packet loss rate < 1%Number of devices: 1000, Packet size: 128B Parameters of the three types of slicing operations. Environmental interaction calculation process (single step Δt = 10ms). Task Generation: # URLLC Task Arrival. arrivals = np.random.poisson(lam = 120*0.01). # ΔtNumber of missions reached within for _ in range(arrivals). task = {“type”: “URLLC”, “size”: 32768, “cpu_req”: 500, “deadline”: current_time + 0.01}; task_queue.append(task). Resource allocation: Intelligent body outputs action vectors , Mapping to physical resources. SLA Default Detection: URLLC: ; eMBB: . Detailed analysis of the calculation of performance indicators System utility function calculation The global utility is a weighted sum with weights set based on service priorities (URLLC:0.6, eMBB:0.3, mMTC:0.1): , where the single-slice utility calculation process (in the case of URLLC). def calc_urllc_utility(slice): latency_violation = sum(task[“latency”] > 0.01 for task in slice.tasks) / len(slice.tasks); resource_cost = 0.05 * slice.allocated_cpu + 0.03 * slice.allocated_bw; return 0.7 * (1 - latency_violation) - resource_cost # Basic utility formula. SLA default rate statistics Default detection algorithm in the case of URLLC: violated_count = 0. for task in completed_urllc_tasks: if task[“start_time”] is None # Not implemented owing to insufficient resources. violated_count + = 1; else if (task[“finish_time”] - task[“arrival_time”]) > 0.01. violated_count + = 1; violation_rate = violated_count / total_urllc_tasks. Calculation of resource utilisation CPU utilisation is obtained through kernel-level monitoring: # Linux perf Tool sampling (every 100ms). perf stat -e cycles; instructions -a sleep 0.1; The mathematical expression is . In-depth analysis of experimental results The statistical results of the mean across 30 experiments are presented in Table 5. MARL-GTRO demonstrated significant superiority over the three benchmark algorithms, Nash Bargaining, PPO (Single Agent) and Round-Robin, in core metrics including system utility, SLA breach rate, and CPU/bandwidth utilisation. Specifically, system utility reached 102.3, representing a 33.2% improvement over Nash Bargaining. The URLLC SLA breach rate was merely 2.7%, a 78.0% reduction compared to Nash Bargaining, fully validating the framework’s resource allocation optimisation capabilities. Table 5Breakdown of 30 experimental means.ArithmeticSystem utilitySLA default rate (per cent)CPU utilization (%)Bandwidth utilization (%)MARL-GTRO102.3 ± 2.1URLLC: 2.7 ± 0.8eMBB: 1.9 ± 0.592.4 ± 1.388.7 ± 2.0Nash Bargaining76.8 ± 5.3URLLC: 12.3 ± 3.1eMBB: 8.2 ± 2.478.2 ± 5.272.5 ± 5.1PPO(Single Agent)60.1 ± 7.8URLLC: 20.6 ± 6.2eMBB: 15.8 ± 5.765.3 ± 6.961.4 ± 7.3Round-Robin45.5 ± 9.1URLLC: 35.7 ± 8.9eMBB: 28.1 ± 7.553.8 ± 8.349.2 ± 9.0 Breakdown of 30 experimental means. URLLC: 2.7 ± 0.8 eMBB: 1.9 ± 0.5 URLLC: 12.3 ± 3.1 eMBB: 8.2 ± 2.4 URLLC: 20.6 ± 6.2 eMBB: 15.8 ± 5.7 URLLC: 35.7 ± 8.9 eMBB: 28.1 ± 7.5 Key Performance Comparison Calculated Utility Improvement in Process: URLLC Default Rate Reduced: The results of dynamic stability verification and quantitative analysis of communication overhead are presented in Table 6. Table 6Dynamic stability validation (standard deviation comparison) and quantitative analysis of communications overhead.IndicatorMARL-GTRONash BargainingStability ImprovementStandard deviation of system utility2.15.360.4%Standard deviation of URLLC default rate0.83.175.2%MethodsSignalling typeSingle-step transmission volumeTotal bandwidth usageCentralised PPOFull state upload3.2 KB/step45.6 MbpsMARL-GTROAction Vector96 B/step15.3 Mbpscomputational process: Dynamic stability validation (standard deviation comparison) and quantitative analysis of communications overhead. Conclusions The experimental results show that MARL-GTRO achieves Pareto optimisation of resource allocation in dynamic MEC environments through the fusion of hierarchical attention mechanism and evolutionary game dynamics, taking into account both SLA guarantee and system stability, and provides a verifiable technological path for intelligent management of 6G network slices.The experiment verified the superiority of MARL-GTRO in three major dimensions. Optimality of resource allocation The system utility reaches 102.3, which is significantly higher than the game theory approach (+ 33.2%) and the single-intelligent body PPO (+ 70.2%), and the key mechanism, VDN-value decomposition, allows the intelligences to pass the Implicit coordination to avoid local optima . SLA guarantees reliability URLLC default rate pressed down to 2.7% (98.1% compliance rate for scenarios with industry requirements < 1%) in the reward functionReduction of resource mismatches and increase in the rate of completion of tasks to 97.3 per cent. System stability breakthrough Under the harsh scenario of ± 30% traffic fluctuation, the standard deviation of utility is only 2.1 (60.4% lower than that of the gaming method), and local decision-making avoids the bottleneck of the central node, and the decision-making latency is stable at 0.84ms (meeting the 5G TTI ≤ 1ms). Communication efficiency innovation 15.3 Mbps signaling overhead 66.4% lower than centralized RL, action space compression (3-dimensional vectors) and observation localization (11 dimensions only). The MARL-GTRO framework achieves Pareto-optimal resource allocation in a dynamic MEC environment, and its Nash equilibrium convergence property makes the four major indicators of system utility, SLA guarantee, stability, and communication efficiency break through the bottleneck of the existing technology, providing a new paradigm for 6G network slice management. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Author contributions This article was completed by Professor Kai Mao . Firstly, performed the data analysis; secondly, performed the formal analysis; thirdly, performed the validation and finally wrote the manuscript. Data availability Due to the nature of this study, which involves proprietary network configurations, sensitive operational data from industry partners, etc., and user data privacy issues, the simulation/experimental data supporting this study is not publicly available. However, aggregated results sufficient to reproduce the main findings are presented. Requests for specific non-sensitive data subsets may be considered by the corresponding authors upon reasonable request and confidentiality agreements. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement Due to the nature of this study, which involves proprietary network configurations, sensitive operational data from industry partners, etc., and user data privacy issues, the simulation/experimental data supporting this study is not publicly available. However, aggregated results sufficient to reproduce the main findings are presented. Requests for specific non-sensitive data subsets may be considered by the corresponding authors upon reasonable request and confidentiality agreements.