<ul><li>Compounding error, where small prediction mistakes accumulate over time, presents a major challenge in learning-based control.</li><li>Mitigating compounding error is important in model-based reinforcement learning and imitation learning.</li><li>Training multi-step predictors directly can help reduce compounding error and improve performance.</li><li>In the context of linear dynamical systems, well-specified single-step models achieve lower asymptotic prediction error, while direct multi-step predictors perform better in case of misspecified models with partial observability.</li></ul>

Learning with Imperfect Models: When Multi-step Prediction Mitigates Compounding Error

Discover more