Q-FOX: A Reinforcement Learning Framework with FOX-Inspired Adaptive Hyperparameter Optimization

Abstract

Translate Article

This paper presents Q-FOX, a novel reinforcement learning (RL) framework that integrates the FOX optimization algorithm into Q-learning to enable adaptive hyperparameter tuning. By dynamically adjusting critical parameters such as the learning rate and discount factor, Q-FOX addresses the limitations of manually tuned Q-learning, which often suffers from suboptimal convergence and unstable policies. The FOX algorithm, inspired by red fox hunting behavior, balances exploration and exploitation in the hyperparameter space. Q-FOX is evaluated on both discrete (GridWorld, Cliff Walking) and continuous (Mountain Car) environments using tabular Q-learning with state discretization. Results demonstrate significant improvements in average reward, success rate, and convergence speed compared to conventional Q-learning, validating the effectiveness of FOX-based tuning. A comparative analysis is conducted to highlight Q-FOX’s robustness under constrained training episodes. The proposed framework shows promise for broader application in robotics, autonomous navigation, and adaptive control systems, where dynamic learning strategies are essential.