Is there any successful implementation of Deep Reinforcement Learning agents in MESA ?

[mrak31]

Hi, I'm wondering if there is any successful implementation of Deep Reinforcement learning agents in MESA.

There are various Reinforcement learning (RL) libraries / baselines to choose from i.e. (https://blog.dataiku.com/on-choosing-a-deep-reinforcement-learning-library).

Is it worth the effort to implement deep reinforcement learning agents in a MESA based models (simulation environments) ? Eventhough GymAi environments are available - although i noticed the focus of GymAi simulation environment (model) are mostly on single agents.

[rht]

I think it is straightforward to implement any Mesa model as a gym environment that can be plugged in with any DRL libraries you use, like RLib. You just need to expose your model step() function. No extra setup need to be done from Mesa (the library)'s end.

[mrak31]

Hoping there was a specific project someone can point to, but thanks

[tpike3]

I can't give you a specific repo but here is a list of repos that use Mesa there maybe one in there https://github.com/projectmesa/mesa/network/dependents

If you do find please post

[mrak31]

Looking into this one:
https://github.com/eescriba/smart-cities-drl

Looks related to what i'm looking for

[jackiekazil]

@mrak31 Thank you for highlighting it! Now I am looking into it!

[rht]

I just discovered whynot.

[jackiekazil]

Can you provide more details as to what you are thinking?

[rht]

Upon reading further, it supports reinforcement learning only for the existing OpenAI gym environments. I thought it had wrapper for Mesa models as an environment for gym. Or perhaps @millerjohnp would know better.