A Context-Partitioned Stochastic Modeling System with Causally Informed Context Management and Model Induction

We describe a flexible multi-layer architecture for contextsensitive stochastic modeling. The architecture incorporates a high performance stochastic modeling core based on a recursive form of probabilistic logic. On top of this modeling core, causal representations and reasoning direct a long-term incremental learning process that produces a context-partitioned library of stochastic models. The failure-driven learning procedure for expanding and refining the model library employs a combination of abductive inference together with EM model induction to construct new models when current models no longer perform acceptably. The system uses a causal finite state machine representation to control on-line model switching and model adaptation along with embedded learning. Our system is designed to support operational deployment in real-time monitoring, diagnostic, prognostic, and decision support applications. In this paper we describe the basic multi-layer architecture along with new learning algorithms inspired by developmental learning theory.