Work in progress...
GymmeForce provides a framework for rapid experimentation with popular Deep Reinforcement Learning algorithms, it focus on making very easy to implement new ideias, which can be rapidly evaluated using OpenAI Gym.
git clone https://github.com/lgvaz/gymmeforce.git
cd gymmeforce
pip install -e . Examples on how to run different agents can be found on the examples folder.
- DQN and extensions
- Double Q-learning
- Dueling networks
- N-step learning
- Soft target update
- PPO
- Clipped Surrogate Objective
- Adaptive KL Penalty Coefficient
- Vanilla Policy Gradient
- REINFORCE
- Actor-Critic
-
DQN on BreakoutNoFrameskip-v4
Episode 0 ---------------- Episode 3500 ----------- Episode 6000 ----------- Episode 7500 ----------- Episode 21500
Mean reward after training: 421 (Averaged over 100 episodes)
Dark blue: Standard DQN
Light blue: Double DQN
-
PPO on Hopper-v1 Video
