An Example of Synthetic Data Generation for Control Systems Using Generative Adversarial Networks: Zermelo Minimum-Time Navigation
Nachiket U. Bapat, Randy Clinton Paffenroth, Raghvendra V. Cowlagi · 2024
Real-world data of the operation of control systems is often scarce because experiments are expensive and time-consuming. We address the problem of synthetic data generation, namely, the problem of generating data about the operation of a physical system through computational means. In this paper, we report generative adversarial network (GAN) models that can incorporate training data and governing equations underlying the operation of a control system. Instead of over-generalization, we restrict the discussion to the particular example, namely, the Zermelo navigation problem of a vehicle navigating a drift field (e.g., wind) in minimum time. Owing to the nature of the example chosen, we find algebraic governing equations. We propose GAN models that learn to generate trajectories that not only resemble the training data but also satisfy these governing equations. We compare the proposed models to a standard GAN model that uses training data, only, i.e., neglects the governing equations, to demonstrate that the proposed models outperform the standard model on several metrics. To the best of our knowledge, this is the first work on generative models for control systems.