The basic premise underlying control is that the true dynamics of the controlled system are faithful to the dynamics used during controller synthesis, whether that synthesis be feedback design or optimal trajectory design. Faithfulness to the dynamics includes both the natural dynamics of the unforced systems and the external forcing applied to the system through environmental interaction. When either or both deviate from reality, then control performance is compromised. Stability guarantees are lost and boundedness results in the bast case (instability in the worst case).
Our research efforts seek to create an online learning strategy for control systems that can estimate the defect between the idealized model and the actual system, then incorporate it into the feedback design or the optimal trajectory design. Initial work sought to introduce tools and methods from machine learning to create a data-driven neuro-adaptive controller, which begins with an empty network then builds and reconfigures the network as needed during operation given a fixed network size. The fixed network size is necessary as a means to constrain the computational budget for real-time operation. Later work explore the outer loop trajectory synthesis with refinements of the underlying dynamics informed by online learning of the dynamics defect.