Data-driven control of nonlinear systems: An online sequential approach

While data-driven control has shown its potential for solving complex tasks, current algorithms such as reinforcement learning are still data-intensive and often limited to simulated environments. Model-based learning is a promising approach to reducing the amount of data required in practical implementations, yet it suffers from a critical issue known as model exploitation. In this paper, we present a sequential approach to model-based learning that avoids model exploitation and achieves stable system behaviors during learning with minimal exploration. The advocated control design utilizes estimates of the system's local dynamics to step-by-step improve the control. During the process, when additional data is required, the program pauses the control synthesis to collect data in the surrounding area and updates the model accordingly. The local and sequential nature of this approach is the key component to regulating the system's exploration in the state–action space and, at the same time, avoiding the issue of model exploitation, which are the main challenges in model-based learning control. Through simulated examples and physical experiments, we demonstrate that the proposed approach can quickly learn a desirable control from scratch, with just a small number of trials.