search

LucasAlegre / sumo-rl

Reinforcement Learning environments for Traffic Signal Control with SUMO. Compatible with Gymnasium, PettingZoo, and popular RL libraries.
https://lucasalegre.github.io/sumo-rl
MIT License
698 stars 190 forks source link

DI-engine + sumo-rl demo #69

Open PaParaZz1 opened 2 years ago

PaParaZz1 commented 2 years ago

Hi, I am the developer of DI-engine, we are developing a new DRL platform with various algorithms and all kinds of environments.

It is very nice of your project and I want to ask whether you are willing to prepare a DI-engine + sumo-rl demo together, like the following examples:

Screen Shot 2021-12-14 at 8 11 17 PM
LucasAlegre commented 2 years ago

Hi @PaParaZz1 !

Sure, what do you have in mind?

PaParaZz1 commented 2 years ago

Hi @PaParaZz1 !

Sure, what do you have in mind?

OK, I think we can do some cooperations by the following steps:

  1. A simple demo about DI-engine + sumo-rl, like current dqn_2way-single-intersection.py for SB3, which environment do you suggest is siutable for this demo, and can you offer the results of SB3 like reward curve for comparison.
  2. Fancy video demo, we can design a visually expressive demo to attract more people for AI in traffic control. Do you have some materials or ideas about this?
  3. Application for sumo-rl with other sub-domain RL research, like imitation learning, offline RL, MARL and so on. DI-engine integrates many different domains RL algorithms, so we can explore some interesting topics such as multi-agent cooperation, behavior cloning on expert policy and data on sumo-rl.
PaParaZz1 commented 2 years ago

Can you provide more results(e.g. training curve) on different cases? We are doing some experiments on this repo with DI-engine and want to compare your previous results using SB3 and RLlib.