Variational autoencoder-based robust learning control for nonlinear systems with piecewise disturbances

Abstract

This paper proposes an adaptive reinforcement learning framework based on a Variational Autoencoder (VAE) for robust control of nonlinear systems subject to piecewise disturbances. The framework addresses incomplete state information in partially observable environments by augmenting states with inferred latent disturbance information. Specifically, a variational inference strategy learns disturbance patterns from observable historical trajectories, and a segment boundary detection mechanism accurately identifies environmental switching points. To ensure learning stability and adaptability in piecewise disturbance environments, an uncertainty-aware adaptive mechanism is proposed that dynamically adjusts exploration intensity and update conservatism in reinforcement learning using VAE inference uncertainty. When environmental cognitive confidence is low, the system automatically enhances exploration and adopts more conservative policy updates to ensure learning stability and effectiveness. The effectiveness of the proposed method is demonstrated through cart-pole control simulations.