Is there any examples for multi model system for RL?

Hey!
I am a game developer. I am familiar with couple of Unity ML agents systems like PPO , SAC and other systems. But having trouble with multi tasking systems .

As most of the compex 3d interactive tasks consist of several small tasks, learning complex task needs to be started from small skills . For inctance, in sports like soccer, basketboll, players do start with just learning individual sport skills rather than playing big task playing with basketball or soccer.
First players learn specific skills , run jump ,kick, then try use these skills in real match.

So here brain should consist of several small models rather than large one NN model.

  1. tiny skills(during every day training)
    2 . large skills that integrates tiny skill (during weekly matches)

Is there any multi model RL examples that first starts to learn simlple tasks, then try to solve big task using individual skills.?

What you want is called hierarchical RL. I doubt unity has a built in agent for it but hugging face may have one since they like unity, not sure Models - Hugging Face. The standard task for this is a soccer game simulation, exactly as you mention. This is a paper of a real robot doing this task on my “not got to that yet” list. [2208.01160] Hierarchical Reinforcement Learning for Precise Soccer Shooting Skills using a Quadrupedal Robot

1 Like