A comparative study of reinforcement learning methods for autonomous navigation in static and deterministic environments

Sammanfattning

This thesis compares four reinforcement learning algorithms for autonomous navigation in static and deterministic grid-world environments: Q-learning, Deep Q-Network, REINFORCE and proximal Policy Optimization. The algorithms were tested on five maps with different levels of difficulty and evaluated using average reward, success rate, greedy evaluation reward, path behavior and training time. The results show that Q-learning performed best overall. It reached successful final policies on all maps and had the lowest training time. PPO also performed well but required more computation. REINFORCE solved several maps but was less stable, while DQN had the highest training time and struggled on the largest map. The study shows that simpler methods such as Q-learning can be effective in discrete and predictable navigation environments. It also shows that the best algorithm depends on the size and complexity of the environment.