Maximum Entropy Semi-Supervised Inverse Reinforcement Learning

Name: Maximum Entropy Semi-Supervised Inverse Reinforcement Learning
Uploaded: 2026-10-08T10:51:57.148058+00:00
Duration: 340 s
Description: This work introduces MESSI, a new algorithm for apprenticeship learning that combines Maximum Entropy Inverse Reinforcement Learning with semi-supervised learning principles. It uses a unique pairwise penalty to integrate unsupervised trajectories, helping to resolve ambiguity in MaxEnt-IRL by favoring reward vectors that assign similar rewards to similar expert behaviors.

This work introduces MESSI, a new algorithm for apprenticeship learning that combines Maximum Entropy Inverse Reinforcement Learning with semi-supervised learning principles. It uses a unique pairwise penalty to integrate unsupervised trajectories, helping to resolve ambiguity in MaxEnt-IRL by favoring reward vectors that assign similar rewards to similar expert behaviors.

Open the full episode