Knowledge Fusion-Based Neural Network Control for Uncertain Nonlinear Systems via Deterministic Learning

This article proposes a knowledge fusion neural network (NN) control method based on deterministic learning (DL) for uncertain nonlinear systems. The goal is to explore the knowledge acquisition capability and generalization of the controller in a larger task space, enhancing the learning ability and control performance for complex control tasks. Specifically, the proposed closed-loop knowledge fusion control scheme is divided into the following two categories: online and offline knowledge fusion learning control (KFLC). In the online KFLC phase, a collaborative control strategy is used, incorporating a mechanism to transmit neural update information. This ultimately ensures that NN weights of all active systems converge to a shared optimal value. Second, offline KFLC initially achieves accurate identification of the intrinsic closed-loop dynamics through DL control for each single trajectory. The knowledge is then stored as constant value NNs, and subsequently, the issue of knowledge fusion for multitrajectory closed-loop dynamics is transformed into a least squares (LS) problem. Furthermore, an NN-based learning controller utilizing integrated knowledge is constructed to achieve the vision of multitask intelligent control in complex scenarios. The simulation section validates the effectiveness of the proposed scheme.