Skip to content

README.md

Latest commit

 

History

History
27 lines (21 loc) · 1.29 KB

README.md

File metadata and controls

27 lines (21 loc) · 1.29 KB

reinfo-learning

Deep Q-Network (DQN) for Continuous Maze Navigation

This Academic project implements a Deep Q-Network (DQN) to solve a continuous 2D maze navigation problem with complex reward structures and obstacles. Through a series of experiments—ranging from step-limit tuning to epsilon decay strategies and architectural improvements—the agent successfully learns both single and multi-path solutions to reach the goal, avoiding danger zones and optimizing rewards.

Environment

  • State Space: 2D continuous position (x, y)
  • Action Space: 4 discrete actions (up, down, left, right)
  • Goal: Navigate from start to goal
  • Obstacles: Walls and danger zones (red rectangles)
  • Rewards:
    • Goal reached: +100
    • Danger zone: -10 (episode ends)
    • Wall collision: -1
    • Distance-based shaping: reward for moving closer to goal
    • Step penalty: -0.01 per step

Multi-Path Learning with Improved Architecture

  • max_steps: 200
  • epsilon: Exponential decay from 1.0 to 0.1
  • Network: 4-layer architecture (256→256→128→4) with dropout
  • Learning rate: 0.001 with gradient clipping
  • Key Achievement: Agent learned multiple optimal paths to the goal
  • Improvement: Enhanced exploration-exploitation balance with proper consecutive success tracking