Wanted: Second opinion on my conclusions about creating an ecosystem.

Hi, I am trying to figure out if, and how, ML-agents can help me in my project of creating an ecosystem with both lifelong learning and evolution between generations.

The way that I want the ecosystem to work is that two or more species with several individuals each should be able to take actions in a world with static objects, e.g. collect food and water, as well as interact with each other, e.g. hunting prey (other species) and mating. I have, however, figured that some limitations might hinder this and I would like to have a second opinion to ensure that I have understood things right.

So, creating the world with the agents in Unity should not be any problem. But to train them, I figure that the only(?) way to do this is through the gym wrapper as I want to develop custom learning algorithms (that should be adaptable and replaceable). When reading the documentation, I found that the gym wrapper does only support one single agent. I guess that this means that it is not possible to have an environment with more species and many individuals?

The follow-up question is if someone sees another approach to this problem formulation within the ML agents framework?

 

Well you can do the following: You can start training in Unity with Multi-agent and after training and getting a neural network for each you can put all their nnmodels.After the first training session, with the help of the gym wrapper you can train one agent in that environment and others will be pretrained agents. You can iterate this process one by one for every agent type that you want to train with the gym and replace their nn model with the newer one. It will be not the same as training with multagent environment but at least it converge something that you might okay with.