0
$\begingroup$

Consider an MDP where the transition model and the reward function are unknown.

Consider an optimal policy $\pi^*$ generated from this MDP (say by some oracle who does know the transition model and reward function).

How can one infer the reward function and transition model from an optimal policy? For such a question, there may be multiple approaches, so this question is mostly a reference request.

Improve this question
$\endgroup$
2
  • $\begingroup$ Maybe im just stupid, but can't we assume that \pi^* is deterministic, and for each state put only one transition with reward $1$ to itself for the action $\pi^*$ does, and for everything else reward 0? $\endgroup$
    – nir shahar
    Commented Apr 28, 2021 at 5:11
  • $\begingroup$ You wont have any way to distinguish between this MDP and any other MDP for which $\pi^*$ is optimal $\endgroup$
    – nir shahar
    Commented Apr 28, 2021 at 6:12

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.