Inverse Reinforcement Learning (IRL)

[Definition from here] Given:

• States \(s \in \mathcal{ S }\) and actions \(a \in \mathcal{ A }\).
• (Sometimes) Transition probabilities \(p(s'\mid s,a)\).
• Sample trajectories \(\{\tau_i\}\) sampled from running the optimal policy \(\pi^*(\tau)\).

Learn:

• The reward function \(r_\psi(s,a)\) where \(\psi\) represents reward parameters.

Then use the learnt reward funcion to learn \(\pi^*(a\mid s)\).

Some papers mentioned from Lecture linked above:

• (, ).
• (, a).
• (, a).
• (, a).