This work presents two Reinforcement Learning setups: a) Actor Critics Deep Deterministic Policy Gradients (DDPG) and b) Deep Q-Learning (DQN). DDPG's Actor and Critic components consists of a Convolutional and a Full Connected Neuronal Network respectively, the former infers agent's actions and the later assess quality of such predictions. For Deep Q-Learning it was constructed a ConvNet that represents the Q-Value function, mapping states to Q-Values used to select actions greedily with agent-environment further interactions without ever stopping exploration behavior (i.e GLIE). Similar setups were previously applied on 2D-world game playing agents, in this work it was applied to Balloma: a 3D-world android game. Deep Q-Learning performed better than DDPG, however currently further tunning is necessary in order to obtain a more practical video game playing agent.
*Please read capstone_report.pdf for the insigts on what has been developed in this repository.
- Minicap.
- Android Debug Bridge (adb)
- Keras.
- Tensorflow backend
- Deep Determinist Policy Gradients
- Deep Q-Learning
- OpenCV
- Install python 3.7.5.
- Activate developer mode on your android device.
- Plug android device to PC and make sure it is usable by adb
adb devicesshows online device. - Install lib dependencies
pip install -r requirements.txt. - Clone minicap
- Run minicap with
./run.sh autosize(You need ndk-build for this) - Forward requests to minicap with
adb forward tcp:1313 localabstract:minicap - Install Balloma game in your android device.
- Open Balloma and start game's first scene.
- Run
python training.py.
- Currently this project is only compatible with Samsung S8+ device.
It will start inputting actions onto the device infered by the Actor's
ConvNet. Scene is restarted automatically after every episode ends. Also you can use the scripts in plots folder of this repo to see training progress through metrics.