Hello, I pushed some python environments for Multi Agent Reinforcement Learning. Some are single agent version that can be used for algorithm testing. I provide documents for each environment, you can check the corresponding pdf files in each directory. Some environments are like:
OpenCV, swig
Assumption:
Each agent works synchronously.
Member Functions
reset()
reward_list, done = step(action_list)
obs_list = get_obs()
reward_list records the single step reward for each agent, it should be a list like [reward1, reward2,......]. The length should be the same as the number of agents. Each element in the list should be a integer.
done True/False, mark when an episode finishes.
action_list records the single step action instruction for each agent, it should be a list like [action1, action2,...]. The length should be the same as the number of agents. Each element in the list should be a non-negative integer.
obs_list records the single step observation for each agent, it should be a list like [obs1, obs2,...]. The length should be the same as the number of agents. Each element in the list can be any form of data, but should be in same dimension, usually a list of variables or an image.
Typical Monte Carlo Procedures
reset environment by calling reset() get initial observation get_obs() for i in range(max_MC_iter): get action_list from controller apply action by step() record returned reward list record new observation by get_obs()