Reference

Why relevant

Sits adjacent to bajgar-2024-valuewalk and the thesis-hub continuous-action extension work — another framework that goes beyond a deterministic point-estimate reward, but takes a different route than ValueWalk’s Bayesian posterior-over-rewards approach.

Summary

  • Proposes DistIRL, an offline IRL framework that jointly models (a) uncertainty over the reward function and (b) the full distribution of returns — not just a deterministic reward estimate or expected-return matching.
  • Core mechanism: minimizes first-order stochastic dominance (FSD) violations, and integrates distortion risk measures (DRMs) into policy learning → recovers both a reward distribution and a distribution-aware (risk-sensitive) policy.
  • Theoretical: convergence with O(ε⁻²) iteration complexity.
  • Validated on synthetic benchmarks, real-world neurobehavioral data, and MuJoCo control tasks; claims SOTA imitation performance.
  • Explicitly does not claim unique ground-truth reward recovery — recovers a reward distribution compatible with the chosen prior, variational family, and FSD-based inductive bias. Same epistemic humility framing as Bayesian IRL / ValueWalk.
  • BIRL framework ≈ their Dis-TD-Mean configuration
  • RIZE (Karimi & Ebadzadeh, 2025, arXiv 2502.20089) ≈ Det-Qt-Mean — performs worst in their ablations
  • Det-TD-Mean ≈ ValueDice, but with explicit reward estimation

Open questions / to check

  • How does the FSD/DRM machinery compare to ValueWalk’s MCMC-over-reward-posterior approach — complementary (could stack) or a competing paradigm?
  • Worth reading Karimi & Ebadzadeh 2025 RIZE (arXiv 2502.20089) and Ke et al. 2025 switching-rewards/animal-behavior IRL (arXiv 2501.12633) — both cited as close relatives, and the animal-behavior angle overlaps with ValueWalk’s neurobehavioral motivation.
  • Check if their D4RL / MuJoCo continuous-control benchmarks give a direct comparison point for the continuous-action-extension-directions work.

thesis