My Playground for RL algorithms

Implemented algorithms:

• REINFORCE with baseline
  • pyTorch documentation implementation
  • my implementation
• DQN
  • stable-baseline3 with my custom environment
  • my implementation

These algorithms are run on CartPole and my own coloring environment Performance of my algorithm is compared to that of stable-baselines3 in order to ensure correctness

clone git@github.com:denisergashbaev/rl_playground.git
cd rl_playground
python3 --version # should be 3.6.* something
python3 -m venv venv
source venv/bin/activate
pip install --upgrade pip # we need new version of pip for requirements.txt
pip install -r requirements.txt

Example script running a DQN algorithm on the ColoringEnv:

python rl_playground/algs/dqn/dqn_main_simple.py --size=10x10 --log_every=1 --start_train_ts=100 --gamma=1.0 --seed=1 --learning_rate=1e-4 --max_grad_norm=10 --step_reward=1 --depth_channel_first --stay_inside --with_step_penalty --with_revisit_penalty --with_color_reward

Result (javascript-based rendering not yet in the repo)