Custom Environment Navigation System Using Reinforcement Learning

This paper presents a comparative analysis of four reinforcement learning algorithms—Q-Learning, Dyna-Q, Dyna-Q+, and Priority Sweeping—within a custom grid-based navigation environment. The study aims to evaluate these algorithms in terms of their convergence speed, adaptability, and efficiency in dynamic settings where obstacles and goals can change. The research identifies key challenges, such as ensuring algorithm robustness and balancing exploration versus exploitation in dynamic scenarios. Experimental results demonstrate that while all algorithms achieve navigation, Dyna-Q+ and Priority Sweeping exhibit superior performance in adapting to rapid environmental changes. These findings highlight the strengths of advanced reinforcement learning techniques in addressing dynamic navigation tasks and provide insights into their applicability for real-world scenarios.