Deep RL for Wumpus World

Project Overview

Developed an advanced artificial intelligence system applying Deep Q-Learning (DQL) to solve the classic Wumpus World problem. The project demonstrated the practical application of modern reinforcement learning techniques in complex decision-making environments, implementing a sophisticated neural network architecture that learned to navigate dangerous terrain while pursuing objectives.

Technical Implementation

• Environment Architecture: Engineered a 10x10 grid world environment featuring dynamic elements (gold, Wumpus) and environmental signals (breeze, flash, stench). Incorporated state management with directional sensing.
• Neural Network Design: Implemented a deep neural network (200 -> 18 -> 200 -> 4 neurons) optimized for the action space (forward, left, right, reverse).
• Learning Algorithm (DQL): Developed a comprehensive implementation featuring:
  • Experience replay buffer (50,000 capacity)
  • Epsilon-greedy exploration with dynamic decay
  • Target network architecture
  • Custom reward function balancing risk and exploration

Key Achievements

• Successfully implemented complex DQL architecture.
• Achieved consistent goal-reaching behavior.
• Demonstrated effective risk-reward balancing.
• Developed comprehensive performance visualization system.