Skip to content

Latest commit

 

History

History
89 lines (68 loc) · 2.59 KB

README.md

File metadata and controls

89 lines (68 loc) · 2.59 KB
Raw

Distributed Proximal Policy Optimization

Distributed Proximal Policy Optimization (DPPO) is a new distributed architecture which has several GPU trainers and CPU samplers. The data sampled from these samplers are stored in Redis, and these trainers are sharing their network parameters by share_memory() in Pytorch, this method is much faster than local and global memory.

Main requirements

  • python == 3.9
  • gym==0.24.1
  • gym-microrts==0.3.2
  • mujoco == 2.2.2
  • torch==1.12.0+cu116
  • redis==4.3.4
  • numba==0.55.2

Install

Install all requirements.

pip install -r requirements.txt

Running the code

We include two environments (Mujoco and Microrts) and two distributions (normal and beta).

│  README.md
│  requirements.txt
│  
├─algo_envs
│  │  algo_base.py
│  │  algo_transformer.py
│  │  ppo_microrts_hogwild.py
│  │  ppo_microrts_share.py
│  │  ppo_microrts_share_gae.py
│  │  ppo_mujoco_beta_hogwild.py
│  │  ppo_mujoco_beta_share.py
│  │  ppo_mujoco_beta_share_gae.py
│  │  ppo_mujoco_normal_hogwild.py
│  │  ppo_mujoco_normal_share.py
│  │  ppo_mujoco_normal_share_gae.py
│  │  __init__.py
│          
├─libs  
│      config.py
│      log.py
│      redis_cache.py
│      redis_config.py
│      utils.py
│      __init__.py
│               
└─train_main_local
        board_start.sh
        board_stop.sh
        checker.py
        mps_start.sh
        mps_stop.sh
        sampler.py
        trainer.py
        train_main_local.py
        train_start.sh
        train_stop.sh

You can train them in train_main_local or their own files.

Train example

python train_main_local/train_main_local.py

Train in their respective files

python algo_envs/ppo_mujoco_normal_share.py

Where to modify our algorithms or network structure

You can design your own reinforcement learning through modifying Calculate class (e.g., PPOMujocoNormalShareCalculate), Calculate class is mainly used to calculate gradient loss and update network parameter.

Also, the network structure could be modified in Net class (e.g., PPOMujocoNormalShareNet) which mainly utilized to devise and initialize network structure, output what you want (e.g., state-value of a state, an action you want take and so on).

papers

https://arxiv.org/abs/2301.10919

https://arxiv.org/abs/2301.10920