- navigate to
cd ./python-reinforcement-learning
- start a new virtualenv
pipenv shell
- install all dependencies
pipenv install
- navigate to
cd ./python-reinforcement-learning
- start the virtualenv
pipenv shell
- run script
python3.7 ./src/unity-connector.py
- open unity project
- load RaceTrack scene
- run scene
If your agent made it through the whole track, a ppo_actor.h5 and a ppo_critic.h5 file will be created. The test script will generate a video visualizing how your agent would reacted given real "markku" images (https://markku.ai/).
- check that a valid model.h5 file is available
- navigate to
cd ./python-reinforcement-learning
- start the virtualenv
pipenv shell
- run script
python3.7 ./src/test.py
- build the game using unity (name of the app in my case is "game-build")
- open comand line and navigate to
cd <path-to-build>/game-build.app/Contents/MacOS
- run the game in batch mode
./game-build -batchmode
- navigate to
cd ./python-reinforcement-learning
- start the virtualenv
pipenv shell
- run the recording script
python ./src/recording.py
- open the unity project and start the game
- the recoding will be saved as soon as the goal is reached or you are off track
- give higher rewards to agents which complete the track with the least steps
- add additional exploraton strategies such as RND or self-supervised prediction
- https://towardsdatascience.com/explained-curiosity-driven-learning-in-rl-exploration-by-random-network-distillation-72b18e69eb1b
- https://medium.com/data-from-the-trenches/curiosity-driven-learning-through-next-state-prediction-f7f4e2f592fa
- navigate to
cd ./python-reinforcement-learning
- start the virtualenv
pipenv shell
- run
tensorboard --logdir ./logs
- start agent training
- open web interface
http://localhost:6006/