Autonomous and Adaptive Systems - Overcooked

This project was developed for the Autonomous and Adaptive Systems course at the University of Bologna. It explores training cooperative agents for the Overcooked-AI environment using Multi-Agent Proximal Policy Optimization (MAPPO). The primary focus is on evaluating the agents' ability to generalize across different game layouts and adapt to partners with varying behaviors.

Setup

after cloning the repository, install the requirements:

pip install -r requirements.txt

Training Agents

To train an agent on one or more layouts, run the following command:

python src/train.py

Command-line Flags

• --layouts: A list of layouts to train on (default: ["cramped_room"]).
• --episodes: The number of episodes to train the agent (default: 1000).
• --random_prob: The probability of using a random agent for each episode (default: 0.0).
• --model_name: The name of the model to save (if not specified a default name will be used based on the layouts).

Testing Agents

To test the trained agents, you can run the following command:

python src/test.py

Command-line Flags

• --agents: A list of agents to test.
• --layouts: A list of layouts to test on.
• render: If set, an episode will be rendered in a GUI window after testing is completed.
• verbose: If set, the results of each episode will be printed to the console.
• --no-test: If set, only renders the trajectories without testing the agents.

Demo

The train_agents.ipynb notebook contains the code used to train and test all agents featured in the experiments. For a demonstration of some of these agents in action, refer to demo.ipynb.