Decision-Focused Learning in MDPs: An Occupancy Measure Approach

Name: Decision-Focused Learning in MDPs: An Occupancy Measure Approach
Uploaded: 2026-10-19T11:01:23.625881+00:00
Duration: 410 s
Description: The authors reformulate decision-focused learning in Markov Decision Processes (MDPs) using an occupancy measure-based linear program (LP) layer. This approach yields a closed-form gradient and scales to large state spaces via learnable aggregation, outperforming existing KKT-based methods in terms of regret and computational cost.

The authors reformulate decision-focused learning in Markov Decision Processes (MDPs) using an occupancy measure-based linear program (LP) layer. This approach yields a closed-form gradient and scales to large state spaces via learnable aggregation, outperforming existing KKT-based methods in terms of regret and computational cost.

Open the full episode