Softstar: heuristic-guided probabilistic inference

摘要: Recent machine learning methods for sequential behavior prediction estimate the motives of rather than itself. This higher-level abstraction improves generalization in different settings, but computing predictions often becomes intractable large decision spaces. We propose Softstar algorithm, a softened heuristic-guided search technique maximum entropy inverse optimal control model behavior. approach supports probabilistic with bounded approximation error at significantly reduced computational cost when compared to sampling based methods. present analyze guarantees, and compare performance simulation-based inference on two distinct complex tasks.