LIBRARIES Learning Augmented Recursive Estimation for Uncertain Nonlinear Dynamic Systems

This thesis addresses the joint problems of state estimation and system identification for partially unknown dynamic systems. We focus on the case where the unknown dynamics are nonlinear time-invariant function of the state. We use a spatially localized learning system to construct a state-dependent map-ping of the unknown nonlinearities. This map then serves as a source of evolving long-term memory for use in tasks of estimation and identification. Standard filtering algorithms (EKFs) that exploit the knowledge stored in the learning system are used to estimate both the system state and the correction terms for the learned mapping. These state/correction term pairings are then used to refine the learned mapping of the unknown dynamics. This approach contrasts with many existing identification techniques that treat the unknown nonlinear-ities as time-varying quantities. By not storing any information about previous estimates of the unknown dynamics, these techniques fail to improve their per