Model-Based Reinforcement Learning via Stochastic Hybrid Models

Optimal control of general nonlinear systems is a central challenge in automation. Enabled by powerful function approximators, data-driven approaches to have recently successfully tackled challenging applications. However, such methods often obscure the structure dynamics and behind black-box over-parameterized representations, thus limiting our ability understand closed-loop behavior. This article adopts hybrid-system view modeling that lends an explicit hierarchical problem breaks down complex into simpler localized units. We consider sequence paradigm captures temporal data derive expectation-maximization (EM) algorithm automatically decomposes stochastic piecewise affine models with transition boundaries. Furthermore, we show these time-series naturally admit extension use extract local polynomial feedback controllers from experts via behavioral cloning. Finally, introduce novel hybrid relative entropy policy search (Hb-REPS) technique incorporates nature optimizes set time-invariant derived approximation global state-value function.