An implementation of MADDPG on grid world navigation use multiprocess, and there is worker for sampling, learner for learning policy, and evaluator for evaluate policy at a fixed interval.
├── Env #
| ├── env.py # base env gird world
├── Core training core#
| ├── HER.py # Hindsight experience replay, actually not use
| ├── actor.py # actor node for sampling data
| ├── evaluator.py # evaluate policy node
| ├── learner.py # learning process node
| ├── logger.py # record necessary info
| ├── model.py # define NN model
| ├── normalizer.py # normalize sampling data
| ├── util.py # some utils
├── arguments #train arguments
├── collection_experiments.py # simple visualize, may use for collecting samples, but not use till now.
├── origin_obstacle_states.txt # make obstacle same at all evironment, for simplifing problem
├── plot.py #plot
├── rollout_test.py # rollout saved policy
├── train.py #training entrance
python train.py # modify args in arguments.py
