Do the examples of reinforcement learning such as the hide and seek learn to solve anything but the environment they start in?

Question

As pictured here.

https://www.youtube.com/watch?v=kopoLzvh5jY

OpenAI implements a study for reinforcement learning with multiple learning agents which appear adversarial. "Millions of rounds" they say allowed these players to learn behaviors. The question is, are they really learning ? If once the algorithm they used was finished and the environment changed would the encoded data and decisions the algorithm learns apply and replicate what was 'learned' in any other virtual environment with the same overall rules ?

Answer 1

You can use the code available here, especially now that Mujoco simulator is free, and you can try it for yourself! From what I have seen, agents learn some of these behaviors described in the paper but lots of the time they behave chaotic. I do not think the agents perform well or solve the problem. Mainly, the study focuses on the emergence of some interesting behaviors and not on performance based on some score.

The agents were trained in random versions of the hide&seek environment so they are able to generalize within the same environment. They do not use pixels as inputs and instead they use coordinates, locations, velocities etc. This means that if you design your own environment with mujoco-world generator and have almost everything the same except the environment configuration you could possibly test how well the model generalizes.