Introduction

It is the github repo for the paper: NerveNet: Learning Structured Policy with Graph Neural Networks.

Dependency

The repo is written in Python 2.7. You might need to modify the code repo for compatibility in Python 3.x. Sorry for the inconvenience!

1. tensorflow >= 1.0.1

pip install tensorflow-gpu

GPU version is not mandatory, since in the current repo, gpu is not used by default.

2. gym >= 0.7.4

gym dependency

apt-get install -y python-numpy python-dev cmake zlib1g-dev libjpeg-dev xvfb libav-tools xorg-dev python-opengl libboost-all-dev libsdl2-dev swig

gym installation via pip

pip install 'gym[mujoco]'

To use the mujoco, we actually need to use the mjkey.txt

3. mujoco

pip install mujoco-py==0.5.7

Note that currently, we only support MJPro 1.31. Please install mujoco 1.31 from the official website, and use the mujoco-py version 0.5.7.

4. Misc

pip six beautifulsoup4 termcolor num2words

Run the code

To run the code, first cd into the 'tool' directory. We provide three examples below (The checkpoint files are already included in the repo):

To test the transfer learning result of MLPAA from centipedeSix to centipedeEight:

python main.py --task CentipedeEight-v1 --use_gnn_as_policy 0 --num_threads 4 --ckpt_name ../checkpoint/centipede/fc/6 --mlp_raw_transfer 1 --transfer_env CentipedeSix2CentipedeEight  --test 100

You should get the average reward around 20. If you want to test the performance of pretrained models, you should use:

python main.py --task CentipedeSix-v1 --use_gnn_as_policy 0 --num_threads 4 --ckpt_name ../checkpoint/centipede/fc/6 --mlp_raw_transfer 1  --test 100

The performance of the pretrained model of MLPAA is around 2755.

Similarly, to get the transfer learning result of NerveNet from centipedeSix to centipedeEight:

python main.py --task CentipedeEight-v1 --use_gnn_as_policy 1 --num_threads 4 --gnn_embedding_option noninput_shared --root_connection_option nN,Rn,uE --gnn_node_option nG,nB --ckpt_name ../checkpoint/centipede/gnn/6 --transfer_env CentipedeSix2CentipedeEight --test 100

The reward of NerveNet should be around 1600. And to test the pretrained model:

python main.py --task CentipedeSix-v1 --use_gnn_as_policy 1 --num_threads 4 --gnn_embedding_option noninput_shared --root_connection_option nN,Rn,uE --gnn_node_option nG,nB --ckpt_name ../checkpoint/centipede/gnn/6 --test 100

The reward for NerveNet pretrained model is around: 2477

To train an agent from sratch using NerveNet, you could use the following code:

python main.py --task ReacherOne-v1 --use_gnn_as_policy 1 --network_shape 64,64 --lr 0.0003 --num_threads 4 --lr_schedule adaptive --max_timesteps 1000000 --use_gnn_as_value 0 --gnn_embedding_option noninput_shared --root_connection_option nN,Rn,uE --gnn_node_option nG,nB

WilsonWangTHU/NerveNet