SofDes Spring'15 final project by Kelly Brennan, Stephanie Northway, and Pinar Demetci. The project aims to use AI and machine learning strategies for pathfinding in PacMan.
What is PacMan-AI?
It's a class project we took on to teach ourselves about AI, specifically reinforcement learning, and even more specifically Q-learning. It's a Pacman game that plays itself and learns from its experience. Each time it makes a move, it updates the weights associated with different features (like how far it is from a ghost or food pellet) based on the expected rewards (both immediate and cumulative) of making that move. At the end of a game, it stores these weights and uses them in the next game. Ideally it will converge on the optimal set of weights.
How to use this project:
Clone the repo:
$ git clone https://github.com/pdemetci/PacMan-AI.git
Run pacman.py:
$ cd PacMan-AI/PacMan1
$ python pacman.py --pacman SimpleQPacman --agentArgs fromPickle=True
There are many other options for pacman.py. Run $ python pacman.py --help for more information.
If you want the weights to initialize randomly, you can set fromPickle=False and see how Pacman fares.
If you want to step through the game one move at a time, you can set --frameTime -1, or any negative number. The default frameTime is 0.1, so you can make it run faster or slower.