Repository for a deep reinforcement learning agent that can play the first level of Donkey Kong Country
This is a final project for COMP4107 Neural Networks course at Carleton University.
- Josh Challenger
- Hadi Cheaito
- Nabeel Warsalee
After lots of training and experimentation, we managed to find a successful model that could complete the first level.
The recording of the run can be seen here
This program allows for the training, testing and replay of models for Donkey Kong Country.
The implementation leverages the Proximal-Policy Optimization (PPO) RL algorithm for training our agent.
- OpenAI's Gym and Gym-Retro APIs
- Stable-Baselines3
Run a training episode by providing a name to save the model under and the number of training steps. Will produce a useable model stored in a zip file, with tensorboard logs produced in a separate log directory to monitor training performance.
To tune hyper-parameters, edit the values of the hyper dictionary in the main.py. For editing the reward weights, change the values of the scenario.json file in the /custom_integrations directory. For more information on gym-retro rewards and variables, see the documentation.
python main.py -n my_model -s 10000
To test a recently trained model, simply provide the name of the trained model and use the testing flag. Testing mode will have the agent play 10 times, and will record the key presses of model that made it the furthest in those runs.
python main.py -n my_model -t
Replay the best recorded run during the test mode by using the replay flag.
python main.py -n my_model -p