An active learning method for inverse reinforcement learning that infers an unknown reward from expert demonstrations while minimizing costly human queries by requesting whole expert trajectories rather than isolated single-state actions. In a Bayesian setting with a prior over rewards and a Boltzmann-rational expert, it adapts the expected-information-gain (EIG) acquisition function to select the environment setup whose demonstration most reduces the entropy of the reward posterior. Since EIG is analytically intractable, it is approximated with a nested Monte Carlo estimator and made efficient via Bayesian optimization with an upper-confidence-bound rule that adaptively allocates sampling budget. In gridworlds the method beats random sampling, maximum Q-value entropy, single-state querying, and a prior trajectory-level heuristic.
This NeurIPS 2024 workshop paper develops an information-theoretic approach to active inverse reinforcement learning (IRL), where the goal is to infer an unknown reward function from expert demonstra…