machinelearning

This repository contains my “Machine Learning” course final project.

Overview

The project explores the application of Reinforcement Learning and Deep Q-Learning in a custom game environment called "Cat-Mouse-Cheese". The goal is for the mouse agent to reach the cheese while avoiding the cat, learning optimal movement strategies across different grid configurations. The development involved both classic Q-learning and Deep Q-Networks (DQN) for comparison on various map complexities.

Contents

• 📁 DQN/ – Contains the implementation of the 10x10 grid with obstacles using Deep Q-Learning. This version uses a neural network to approximate Q-values, allowing the agent to learn in a more complex and continuous state space.

• 📁 RL Q Tables/ – Core project using classic Q-learning with Q-tables:

  • 📁 principale/ – The main implementation built collaboratively by the group. It includes:

    • 5x5 empty grid
    • 5x5 grid with walls
    • 10x10 grid with obstacles These environments implement a high-dimensional Q-table, allowing fine-grained control and evaluation of the agent’s learning across scenarios.

  • 📁 alternativo/ – An alternative version developed by one group member. It features:

    • Custom 10x10 grid with "deadly trap" mechanics
    • Variants with and without the presence of the cat This version explores experimental changes to the environment and reward logic.

• 📁 esposizione/ – Files sourced from a recommended GitHub repository (as suggested by the professor). These were used for classroom presentation and not written by me.

• 📄 Relazione Machine Learning.pdf – The final report documenting the complete development cycle, including:

  • Theoretical background on Reinforcement Learning and Q-learning
  • Description of all implemented environments
  • Reward design and training strategy
  • Graphical visualization and performance analysis
  • Comparison between the different learning approaches and environments

Purpose

This repository showcases our collaborative work in building intelligent agents capable of learning optimal strategies through experience. The project focuses on:

• Applying and comparing Q-learning vs. Deep Q-Learning
• Custom environment design with increasing difficulty
• Implementing reward-based learning logic
• Graphical simulation and real-time rendering with Pygame
• Tracking and analyzing performance over time