Evaluating the Efficiency of Reinforcement Learning Algorithms in Dynamic Environment Simulations

Abstract

Reinforcement Learning (RL) has witnessed substantial advancements in recent years, particularly in its application to complex, dynamic environments such as robotics, autonomous driving, and adaptive control systems. This paper presents a comparative evaluation of popular RL algorithms—Q-Learning, Deep Q-Networks (DQN), and Proximal Policy Optimization (PPO)—within high-variability simulated environments. Using metrics such as cumulative reward, convergence time, and policy stability, we examine algorithmic efficiency under different dynamic transition patterns and stochasticity levels. Our simulations, conducted in using the OpenAI Gym and Unity ML-Agents frameworks, reveal that PPO consistently outperforms other methods in highly dynamic scenarios. The findings underscore the importance of balancing exploration and exploitation when deploying RL in real-world systems characterized by continuous environmental shifts.