Maze Solver with Reinforcement Learning (RL) using Reward Shaping in Unity

This project demonstrates a Maze Solver game built with Reinforcement Learning (RL) using Reward Shaping techniques, developed as part of the Advanced Topics on Intelligent Systems course at FEUP (Faculty of Engineering of the University of Porto). The goal of this project is to solve a maze by training an AI agent using RL, improving its decision-making process through reward shaping to navigate efficiently.

Authors:

• Bruno Costa
• Pedro Oliveira

Final GPA

20/20

Key Features:

• Reinforcement Learning: The AI agent learns optimal movement strategies through trial and error.
• Reward Shaping: A technique used to modify the reward function to speed up learning by guiding the agent towards desirable behaviors.
• Unity Game Engine: Developed using Unity, providing a flexible environment for visualizing and interacting with the maze and the RL agent.
• Cross-Platform: The project includes executable files for Linux, Mac, and Windows systems, ensuring compatibility across different platforms.

Installation & Usage:

1. Download the Executable:

   • You can find the pre-built executables for Linux, Mac, and Windows in the git section of this repository.

2. Run the Game:

   • Simply download the appropriate file for your operating system and run the executable.

3. Game Instructions:

   • The game starts with a main menu that leads to 3 different scenes:
     • Menu:
     • Scene 1:
     • Scene 2:
     • Scene 3:
   • The agent must find the path from the start to the end using reinforcement learning techniques.
   • The agent receives positive or negative rewards based on its actions.
   • You can observe the agent's learning progress and decision-making process over time.

Technologies Used:

• Unity: Game engine for creating and visualizing the maze environment and the RL agent.
• Reinforcement Learning (RL): The core technique for training the agent, allowing it to improve its decision-making over time.
• Reward Shaping: Used to modify the reward function, providing additional guidance to the agent during training.

Development Repository:

• The repository containing the developed code can be found here devel.

Conclusion:

This project showcases how Reinforcement Learning and Reward Shaping can be applied to solve dynamic and complex maze navigation problems. The game provides an interactive way to visualize RL concepts while offering a challenging environment for the AI agent.