flow_rl_project

The goal of this project is to develop an multi-agent for multi-merge env in "flow".

Setup

1. docker pull syuntoku/flowtouch:ray_latest

2. git clone [email protected]:syuntoku14/flow_rl_project.git

3. git clone [email protected]:syuntoku14/flow.git

4. modify "syuntoku14" in flow_rl_project/bash_codes/run_flow.bash as your username.

5. modify the port number in flow_rl_project/bash_codes/run_jupyter.bash

How to use

Go inside the docker container

Run

./flow_rl_project/bash_codes/run_flow.bash

Run jupyter notebook

Run

./rl_project/bash_code/run_jupyter.bash

inside the container

Example codes are in ./rl_project/flow_codes/analyze

Train an agent by rllib

The codes are written in flow/flow/benchmarks/rllib

For example, if you want to train ppo with 64 cpus,

python ppo_runner.py --benchmark_name multi_merge --num_cpus 64

Parameter tuning

If you want to tune the parameter, just modify the runner.py codes.

For example,

config["parameter_noise"] = ray.tune.grid_search([True, False])

in ddpg_runner.py

If you want to modify the scale of the reward, modify flow/flow/multiagent_envs/merge.py