Awesome

rl-painter

Reinforcement Learning to Paint

Brief

Each observation contains two images, the "target" and the "canvas". Given the observation, you should output a list of 8 scalars within the range of [0,1] as the action, to control where to paint the next stroke.

The reward will be positive if you decreased the difference between "target" and "canvas". The higher the total reward, the better your agent have painted.

Usage

python env.py to test the environment.

ipython -i ddpg2.py then r(10000) to test the env with a naive DDPG algorithm.

Dependencies

• opencv-python with OpenCV3.x
• a few other helping libraries. Please refer to code.
• Python 3.5+

RL-specific details

• the observation is Markovian.
• human could achieve this task by trial-and-error before finally making a step. Therefore the optimal policy might involve some other classical algorithmic component(search, local optimization, or even modeling) atop deep neural networks.