A Reinforcement Learning–Inspired Latent Yield-based Adaptive Algorithm Switching Mechanism
Jayprakash S. Nair, Jimson Mathew, Shivashankar B. Nair · 2026
Read A Reinforcement Learning–Inspired Latent Yield-based Adaptive Algorithm Switching Mechanism by Jayprakash S. Nair, Jimson Mathew, Shivashankar B. Nair (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-24436.txt.
Introduction
Jayprakash Nair, Mathew, and Shivashankar Nair propose a latent-yield method for nimble switching. Picking the best method per instance is hard in online settings, where traits shift over time, and judging by instant scores alone gives shaky switching. Their technique folds rewards and fines into a latent yield that triggers use and search, pooling scores across instances. Island models from gene methods let method groups explore side-by-side and exchange score notes. Tests on sorting and robot block dodging show the method works.
Catalog record
| CREATOR | Jayprakash S. Nair, Jimson Mathew, Shivashankar B. Nair |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 37716 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.24436
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Reinforcement Learning–Inspired Latent Yield-based Adaptive Algorithm Switching Mechanism Jayprakash S. Nair, Jimson Mathew, Shivashankar B. Nair Abstract Selecting the most suitable algorithm for a given problem instance remains a challenging task, particularly in online or dynamic environments where problem characteristics evolve over time. Relying solely on instantaneous performance metrics can result in a reactive and unstable behaviour, often leading to suboptimal algorithm switching. This paper introduces a computationally efficient approach for aggregating an algorithm’s performance across multiple problem instances that is fairly immune to erratic variations in instance features. Inspired by features inherent to Reinforcement Learning (RL), this technique encapsulates rewards and penalties into a latent yield that, in turn, triggers exploitation and exploration, consequently resulting in adaptive algorithm switching. The proposed technique employs island models, inspired by Genetic Algorithms, to facilitate parallel exploration and performance exchanges among algorithm populations inhabiting local repertoires. Experimental evaluations on sorting algorithms and robotic obstacle-avoidance tasks demonstrate the feasibility and effectiveness of the approach, highlighting its potential in domains where adaptive algorithm selection is critical. 1 Introduction In the realm of AI, Algorithm Selection (AS) is a hot area that endeavours to choose the near-best algorithm given a set of candidate algorithms that can solve a problem instance [ 1 ] . The problem of AS concerns identifying the most appropriate algorithm for solving a given problem instance or a class of problems. A typical AS comprises of a repertoire of algorithms, a problem instance space, a performance measure space and a feature space, all of which play a vital role in facilitating the switch to another algorithm. Despite the abundance of algorithms for a given computational task, their performance can vary drastically across instances due to factors such as input structure, resource constraints, and problem scale. Selecting an unsuitable algorithm may lead to wasted computation and suboptimal results, whereas a well-chosen algorithm can deliver significant performance improvements in efficiency, accuracy, and robustness. The need for AS arises out of the fact that no single algorithm performs best across all problem instances (No Free Lunch Theorem, NFLT) [ 3 ] . This is especially true in cases where the environment is dynamic, where sticking to a pre-decided algorithm may not be worthwhile. Influenced by developments in machine learning, meta-optimization, and adaptive systems, the field of AS has evolved significantly and, as such, forms a cornerstone of automated machine learning (AutoML), evolutionary computation, and intelligent control systems, enabling computational systems to learn and adapt their behaviours based on empirical evidence. AS can be performed either in an offline or online manner. Most of the work is concentrated on the former method. However, if the mapping of the algorithms does not capture the nature of all types of instances, then poor selections are bound to degrade performance. Offline algorithm selection methods neglect the information regarding the performance of the selected algorithm on new online instances [ 2 ] . AS has been applied in various domains, including Graph Neural Networks (GNNs) [ 4 ] , data-driven models [ 5 ] , and multi-armed bandit algorithms [ 6 ] . It can be used to train an ANN to select the best algorithm. Traditional algorithm selection methods assume static problem conditions, but many real-world systems, such as cloud computing, robotics, and streaming data processing, require online adaptive selection of algorithms. In such systems, the algorithm choice needs to be revised continuously based on feedback from the environment. Reinforcement Learning (RL) and Bayesian optimization have proven useful for such adaptive strategies [ 7 ] . Evolutionary computation and Population-Based Training (PBT) methods evolve algorithm portfolios, share performance information, and co-adapt algorithmic behaviour to dynamic conditions [ 8 ] . These hybrid frameworks become especially relevant when it comes to optimization and autonomous decision-making systems. In this paper, we propose a novel and computationally simple method for online algorithm switching that constantly monitors the performance of the algorithm, accumulates and converts it into a latent yield. This latent yield, which constitutes a memory of the aggregate of the past performances of the algorithm, is used to make a decision on whether or not to perform a switch. While this latent yield is synonymous to a combination of rewards and penalties, just as in a typical RL algorithm, the proposed method also inherently performs exploitation and exploration. While this latent yield is synonymous to a combination of rewards and penalties, the proposed method also inherently performs exploitation and exploration, making it akin to RL. In addition, the method makes use of parallelism by conforming to the island models used in Genetic Algorithms (GA). This makes it scalable, both in terms of parallelism and an increase in the number of algorithms within the respective algorithm repertoires. The significant contributions of the work described in this paper include – 1. A cumulative latent yield that remembers the overall performance of an algorithm 2. Delayed switching, thereby avoiding reflex action 3. Parallelism in AS using island models 4. Scalability in terms of an increase in both the number of islands and new algorithms Subsequent sections describe the main methodology, system dynamics, exploration and exploitation mechanisms, followed by results when the AS was applied to a closed and open world problem. 2 Related Work The algorithm selection problem was first formalized by Rice [ 11 ] , who proposed mapping problem instance features to the most suitable algorithm. Early approaches focused on offline selection , where algorithms were chosen once before execution based on static features or empirical benchmarking. With the advent of machine learning–driven selection, systems such as Satzilla [ 14 ] and ASlib [ 9 ] leveraged supervised learning to predict the best algorithm using feature-based meta-models. However, these methods operate in static or batch settings and are unsuitable for online or continuously evolving environments, where problem distributions shift over time. Recent work has therefore explored adaptive and online AS, with RL and bandit-based methods enabling dynamic switching among algorithms during execution [ 10 ] . While effective, these approaches often rely on instantaneous reward feedback, which can cause instability or premature switching when performance fluctuates rapidly. Island model architectures, originally developed in genetic algorithms [ 13 ] , have also been used to promote diversity in optimization by enabling parallel exploration with periodic migration. Similarly, decentralized and distributed frameworks inspired by biological systems have been proposed for online algorithm selection in networked environments. Semwal [ 12 ] describes a decentralized Cyber–Physical System (dCPS) mechanism that evolves mappings between problems and solutions across a network of heterogeneous nodes. Their system continuously adapts to a stream of varying problems, shares evolved solutions among nodes, and demonstrates real-world applicability in both sorting tasks and robotic path-following. This work highlights the benefits of distributed knowledge sharing and continuous learning for selecting optimal solutions without relying on central coordination. This paper situates itself at the intersection of these paradigms — proposing a latent yield–based cumulative estimator for algorithm performance, combined with an island-based exploratory–exploitative architecture, designed to prevent knee-jerk switching. 3 Methodology In computational science, optimization, and multi-agent systems, island models (also called distributed models or archipelagos) [ 15 ] refer to a way of structuring a population of agents in semi-independent groups called islands.