This repo contains an agent for MultiAction Reinforcement Learning. It can be used when the action is discrete (hence a gym.spaces.MultiDiscrete action space) or when the action is continuous (hence a gym.spaces.Box action space).
If the action is continuous, the Environment Wrapper will dicretize the action space. Each action value will be divided into nb_bin discrete values. It has been proven that this method works quite well compared to conventional continuous action RL methods such as DDPG here. The agent is using Branching DQN to allow Q-learning with multiple actions and can be used on almost any environment. #Installation Please install the dependencies by running the following command with pip:
pip install -r requirements.txt
This agent was used to train the Pendulum-v1 environment. Its results for different number of bins are
It was also trained on Bipedal Walker-v3 environment. Its results for different number of bins are
