Improving Multi-Agent Cooperation using Theory of Mind

Content

• Summary
• Installation
• Gameplay
• Simulation

Summary

This repository contains the code for modeling agents through inverse planning with POMDPs on the agent to infer goals and subsequently predict the actions of the other agents in a cooperative setting.

Installation

Clone this repository:

git clone https://github.com/terryyylim/ToM-gameplaying-POMDP.git

To prevent complications to the global environment variables, I suggest creating a virtual environment for tracking and using the libraries required for the project.

1. Create and activate a virtual environment (run pip3 install virtualenv first if you don't have Python virtualenv installed):

virtualenv -p python3 <desired-path>
source <desired-path>/bin/activate

2. Install the requirements:

pip install -r requirements.txt

Gameplay

To start the game, run the following.

python overcooked_server/game.py

Controls

Human Agent 1

Actions	PICK	CHOP	COOK	SCOOP	SERVE	DROP	STAY
Key-Map	Z	X	C	V	B	N	M

Movement	UP	DOWN	LEFT	RIGHT
Key-Map	UP	DOWN	LEFT	RIGHT

Movement	DIAG-UP-LEFT	DIAG-UP-RIGHT	DIAG-DOWN-LEFT	DIAG-DOWN-RIGHT
Key-Map	,	.	/	RSHIFT

Human Agent 2

Actions	PICK	CHOP	COOK	SCOOP	SERVE	DROP	STAY
Key-Map	1	2	3	4	5	6	7

Movement	UP	DOWN	LEFT	RIGHT
Key-Map	Y	H	G	J

Movement	DIAG-UP-LEFT	DIAG-UP-RIGHT	DIAG-DOWN-LEFT	DIAG-DOWN-RIGHT
Key-Map	Q	W	E	R

Simulation

To run a simulation, run the following with desired parameters in CLI.

python overcooked_server/game.py --num_ai_agents=2 --is_simulation=True --simulation_episodes=500