DQN Pacman

Author: Fabian Mora

An implementation of 4 multi agent learning algorithms for playing the Pacman game in the Pacman Berkeley Simulator.

Relevant implemented files

• agent.py defines the interface for all implemented Pacman agents.
• agentUtil.py contains functions for computing state representations of the game.
• rewards.py contains the internal reward function used by the agents.

Implemented agents:

• phcAgents.py based on the PHC algorithm.
• wphcAgents.py based on the WPHC algorithm.
• dqnAgents.py based on DeepQ networks.
• qdqnAgents.py based on DeepQ networks and a WPHC agent.

Running the agents

The recommend mechanism for running an agent is to use the run.py script. For a list of available options and their description use:

python3 run.py -h        # For general options 
python3 run.py PHC -h    # For PHC options
python3 run.py WPHC -h   # For WPHC options
python3 run.py DQN -h    # For DQN options
python3 run.py WDQN -h   # For WDQN options

Examples

python3 run.py -g 10 -t 2 WPHC           # Run the WPHC agent for 10 evaluation games and 2 training games.
python3 run.py -g 10 -t 100 DQN -L 0.001 # Run the DQN agent for 10 evaluation games and 100 training games using an optimizer learning rate of 0.001.