Maze Solver: Reinforcement Learning with Tabular and Deep Methods

Overview

This project explores the Maze Rider problem, a challenge demonstrating the trade-off between exploration and exploitation in Reinforcement Learning (RL). The implementation covers both tabular RL methods and Deep RL techniques to solve mazes of varying complexities efficiently.

Features

• State Space Reduction: Compact representation focusing on key elements (agent and goal positions).
• Reward Shaping: Methods to improve convergence by penalizing wall hits and repetitive actions.
• Sampling Techniques: Implementation of ε-greedy exploration and Thompson Sampling.
• Tabular RL Algorithms: Q-Learning, SARSA, Double Q-Learning, and Dyna-Q.
• Deep RL Methods: DQN, Reinforce, and A2C with MLP and convolutional neural networks.

---

Algorithms Implemented

Tabular RL

• Q-Learning
• SARSA
• Double Q-Learning
• Dyna-Q

Deep RL

• Deep Q-Networks (DQN)
• Reinforce
• Advantage Actor-Critic (A2C)

Sampling Strategies

• ε-greedy Exploration
• Thompson Sampling

For a detailed analysis of algorithms and their performance, refer to the report.