This is an Inverse Reinforcement Learning (IRL) problem. I have data (observations) on actions taken by a (real) agent. Given this data I want to estimate the likelihood of the observed actions in a Q-learning agent. Rewards are given by a linear function on a parameter, say alpha.

Thus, I want to estimate the alpha that makes the observed actions more likely to be taken by a Q-agent. I read some papers (i.e. Ng & Russel 2004), but I found them rather generalistic.

  • 1
    $\begingroup$ Hi and welcome to this community! This is an interesting post, but what is your question exactly? $\endgroup$ – nbro Jun 4 at 20:59
  • 1
    $\begingroup$ HI, thanks! My question is how could I implement an IRL algorithm, like [Ng and Russell, 2000], to a Markov game where multi-agents have opposite interests. The game is simple though, as it is a matrix game. I read some papers on IRL, but neither of them accounts for a multi-agent framework. Actually what I ask is for any ideas on how to approach to a problem of this nature. $\endgroup$ – Julian Lopez Baasch Jun 5 at 17:52
  • $\begingroup$ Have you looked into the work done by Deepmind on AlphaStar for StarCraft? They use a combination of different methods to achieve their goal - one of which is IRL. You can read more about their outline of implementation here: deepmind.com/blog/article/… $\endgroup$ – Krrrl Oct 27 at 21:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.