Overview
The program is designed for MATLAB hack day. It tries to bridge Matlab to OpenAI Gym platform and potentially Simulink.
Software requirement:
Game Platform: require python3, gym, pygame, gym-ple and osim-rl
- The flappy bird in gym-ple has a bug that incorrectly implement the hit condition. To use the correct version, go to "PyGame-Learning-Environment" folder in this repo, and type "pip install -e ."
- To run the human NMS system from NIPS Learn to run, make sure that you have installed and activated the conda environment as suggested in osim-rl.
- Control: developed in Matlab 2017a and linux environment
To run the code:
Firstly setup server by typing "python gym_http_server.py", then run the "example.m" to test simple/rl controller. Feel free to develop your own controller from AbstractController.
Design Idea:
- The http-api design refers to gym-http-api
- For the flappy bird game, .mat file is used to transmit images to avoid overhead in jsonify
To do:
- Implement DDPG algorithm to address continuous control problem in RL framework
- Use system identification to identify system model and develop optimal controller