README.md

MetaVIM

MetaVIM (Meta Variationally Intrinsic Motivated) is a RL method, aim to learn the decentralized policies of each traffic signal only conditioned on its local observation.

Usage and more information can be found below.

Installation

This repo is tested on Python 3.6+, CityFlow 0.1, PyTorch 1.0.0+ (PyTorch 1.5.1 for examples) and TensorFlow 1.14.0.

Dataset

This repo containes four real-world datasets (Hangzhou, Jinan, New York and Shenzhen), which are stored in the ./data

• Road network map

  • File Name	City
    roadnet_4_4.json	Hangzhou
    roadnet_3_4.json	Jinan
    roadnet_16_3.json	New York
    fuhua_cityflow.json	Shenzhen
    roadnet_3_3.json	grid_road_network
    roadnet_6_6.json	grid_road_network
    roadnet_10_10.json	grid_road_network

• Traffic flow

  • *raw denotes the real data.
  • *2570 denotes the mixed_low data.
  • *4770 denotes the mixed_high data.

Usage

Run the code:

1. Firstly, set the city_and_configuration in main.py according to the following format:

[City]_[Pattern]

• City can be hangzhou, jinan, newyork, shenzhen.
• Pattern can be real, mixed_low, mixed_high.

2. Secondly, run the main.py :

Example1:

CUDA_VISIBLE_DEVICES=1 \
python main.py \
  --log_name 6_6_withneighbor_2570 \
  --road_net 6_6 \
  --volume 300 \
  --suffix 0.3_bi \
  --path_to_log records/test/6_6/train_round \
  --path_to_work_directory data \
  --num_processes 36 \
  --use_neighbor True \
  --use_intrinsic_reward True \
  --decode_state_neighbor True \
  --decode_reward_neighbor True \
  --input_action_neighbor True

Example2:

CUDA_VISIBLE_DEVICES=1 \
python main.py \
  --log_name hangzhou_withneighbor_2570 \
  --road_net 4_4 \
  --volume hangzhou \
  --suffix real \
  --path_to_log records/test/hangzhou/train_round \
  --path_to_work_directory data \
  --num_processes 16 \
  --use_neighbor True \
  --use_intrinsic_reward True \
  --decode_state_neighbor True \
  --decode_reward_neighbor True \
  --input_action_neighbor True