Reinforcement learning algorithms applied to the Pacman game.
The Pacman game is the result of Pushkar's work for the Udacity class ud820.
The learners can be found in the qlearningAgents.py file.
python pacman.py -p PacmanQAgent -x 800 -n 810 -l smallGrid
- The
-lparameter can be changed to use other maps (available in the layout folder). - The
-pparameter can be changed to use other types of q-learners. -xsets the number of training examples.-nsets the total number of examples (train + test)
- PacmanQAgent: basic Q-learning agent. It uses full states (the whole map), and thus doesn't scale when using big maps.
Run it:
python pacman.py -g DirectionalGhost -p PacmanQAgent -x 800 -n 810 -l smallGrid
- ApproximateQAgent: Uses a few predifined features to describe a state. What is learnt is how to combine these features to take a decision. This one can be used on big maps. The main drawback of this method is that Pacman is almost blind, as it only know what's happening in his 4 adjacent cells. It can easily get stuck in a corner because it doesn't know where the food is.
Run it:
python pacman.py -g DirectionalGhost -p ApproximateQAgent -x 400 -n 410 -l smallGrid
- PerceptronQAgent: Learns using a perceptron. The input is what is in the cells near-by. Unlike with the previous learner, Pacman has access to a range of cells around him (defined via parameter). [under development]
Run it:
python pacman.py -g DirectionalGhost -p PerceptronQAgent -x 2000 -n 2010 -d