Q-Learning Grid World Solver

This project implements a Reinforcement Learning agent that learns to navigate a 6×6 grid. The agent must find the optimal path from a starting position to a goal while avoiding static obstacles.

🎮 The Environment

The world is a coordinate-based grid where:

• Grid Size: 6×6

• Start Position (S): (5, 0) (Bottom-Left)

• Goal Position (G): (0, 5) (Top-Right)

• Obstacles (X): Static blocks located at specific coordinates that penalize the agent.

🧠 Reinforcement Learning Logic

The agent uses the Q-Learning algorithm to populate a 3D Q-Table (6, 6, 4), representing 4 possible actions (Up, Right, Down, Left) for every grid cell.

Key Hyperparameters

• Alpha (α): 0.3 (Learning Rate)

• Gamma (γ): 0.95 (Discount Factor)

• Epsilon (ϵ): Starts at 0.9 and decays over time to balance exploration and exploitation.

Reward Structure

• Goal Reach: +100

• Obstacle/Boundary Hit: -10

• Each Step: -0.5 (Encourages the shortest path)

🚀 How to Run

1. Initialize the Q-Table with zeros.

2. Run the train_agent() function for 20,000 episodes.

3. Use visualize_best_grid(q_table) to view the learned policy in the console.

📍 Final Learned Path

Below is the visual representation of the agent's optimal policy after training. Arrows indicate the action with the highest Q-value for each state.

---

🛠 Features

• Epsilon-Greedy Policy: Ensures the agent explores the grid thoroughly before settling on a path.

• Boundary Protection: The is_valid_state function prevents the agent from leaving the 6×6 area.

• Detailed Visualization: Formatted console output using :7.2f for aligned and readable Q-values.