Inverse reinforcement learning (IRL) is a task that can learn a reward from other agent behaviour. Most IRL paradigms assume that dynamics of environment is known, that is the transition probability and agent's policy are given.

However, in my problem, the transition probability and specific policy of the agents are unknown and the only thing I have is a bounch of sample trajectories consisted by state and action sequences. So how can I do IRL on such task ? So far, the only example is the typical linear IRL discussed by AY.Ng (Algorithms for Inverse Reinforcement Learning). However, the solution seems not elegant because we still need to define a set of policies as benchmark, which is not deterministic.

Is there any unified framework to do IRL task with sampled trajectories only ? Non-parametric method might be a better solution.


1 Answer 1


I am also looking into this. Two works I found that might be a lead are:

  1. Structured Classification for Inverse Reinforcement Learning
  2. A Cascaded Supervised Learning Approach to Inverse Reinforcement Learning
  • $\begingroup$ Its just url answer $\endgroup$
    – fuwiak
    Commented Aug 15, 2023 at 9:58

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.