EUTOPIA Pacman Contest

Overview

The EUTOPIA Pacman Contest is a competitive team-based challenge where participants develop intelligent agents for a capture-the-flag variant of Pacman. Based on Berkeley's CS188 course and adapted by UPF with contributions from RMIT and UoM, this Python-based framework provides a platform for implementing and testing advanced AI strategies.

Game Rules

• Map Structure: The playing field is divided into two halves, with each team defending their side while invading the opponent's territory.
• Objective: Collect food pellets from the opponent's side and return them safely to your territory.
• Scoring: Each food pellet returned to your territory earns one point.
• Victory Conditions: Win by collecting all opponent's food or having the highest score when time expires.
• Constraints: Players must manage computational resources efficiently to ensure timely actions.

Repository Content

This repository provides implementations of advanced AI agents for the EUTOPIA Pacman Contest:

Getting Started

1. Clone the EUTOPIA contest repository:

   git clone https://github.com/aig-upf/pacman-eutopia

2. Replace team_name_X with your team's name
3. Add the desired version file (v1.py or v2.py) to your repository

Agent Implementations

Version 1 (v1.py)

• OffensiveQLearningAgent: Employs Q-learning for strategic offensive play with dynamic action selection, feature-based decision making, and Bayesian inference for tracking opponents.
• DefensiveReflexCaptureAgent: Focuses on territory defense using patrol strategies around key chokepoints, with Bayesian belief distributions to track invaders when not directly visible.

Version 2 (v2.py)

• OffensiveQLearningAgent: Enhanced Q-learning agent with weight persistence and adaptive strategy development.
• DefensiveQLearningAgent: Q-learning defender that focuses on protecting territory and intercepting invaders, with patrol strategies around key entry points and Bayesian inference for enemy tracking.

Pre-trained Weights

The repository includes optimized weights for both offensive and defensive agents, enabling immediate high-performance gameplay without training:

• default_offensive_weights.json: Optimized weights for offensive strategies
• defensive_agent_weights_613de35e-a3b5-4f6b-bbe5-579884aa87f0.json: Refined weights for defensive play

Key Features

• Reinforcement Learning: Both versions leverage Q-learning to improve performance over time.
• Bayesian Inference: Used for probabilistic tracking of opponents even when not directly visible.
• Adaptive Strategies: Agents dynamically adjust between offense and defense based on game state.
• Strategic Patrolling: Defensive agents identify and monitor key map locations to intercept invaders efficiently.