- Johan Hammarstedt, jhammarstedt
- Lukas Vordemann, vordemann
To build a team to play CTF- Pacman, created by the competition hosted at Berkley. This was built as a school assignment for the class DD2438 at KTH with an inter-class tournament, which we ended up winning.
Our orignal idea was to train a DQL agent to play capture the flag pacman using CNN + fully connected layers. However, the setup for the agent is there, but due to the short deadline we didn't have time to train it sufficiently and as a result very little learning was done.
As a backup to the DQL agent we also added a heuristic based team, that ended up performing really well and won the final tournament. Setup:
- Defensive Agent:
- Uses A* to get all paths
- Tracks enemy movement by checking where food disapears (since you don't have vision of enemies if they're too far away)
- Rush to capsules at beginning to prevent it from being stolen
- Offensive Agent:
- Find closest pill with A*
- "Risk Assesment", will only eat aroud 5 foods at the time to avoid overextending
- If chased it will compute the best path to base using A*, or check for a closer capsule while avoiding enemies
- A flanking manouver to avoid getting stuck with enemies
Use the following command to run the code for the heuristic vs a baseline team
$ python capture.py -r heuristicTeam -b baseline -l RANDOM50
To train the DQN agent:
$ python capture.py -r DQN_agent -b baseline -n 5000 -x 4000 -l RANDOM99
-nis number of episodes-xis number of trainings-lchoose a random grid
Credits to cshelton for rewriting the enviroment to support python 3, you can find the repository here
- CFT Pacman
- starcraft RL
- RL in pytorch: Mario, Flappy Bird
- Keras RL
- MARL on github
- Pacman DQL, repo