IMFlow: inverse modeling with conditional normalizing flows for data-driven model predictive control
- Inverse modeling is the process uncovering the relationships from the system observations to its inputs. It is essential in various fields such as control, robotics, and signal processing. We propose an inverse modeling method using amortized variational inference based on conditional normalizing flows (IMFlow). IMFlow is data-driven and can therefore be applied to black-box environments with limited observability and unknown complexity. Besides, the probabilistic modeling characteristics of conditional normalizing flows allow IMFlow to cope with unknown system uncertainties. We deploy IMFlow as a probabilistic model predictive controller, which estimates the control inputs as stochastic processes based on reference signals and system responses. In addition, we also adjust IMFlow to an online model-free reinforcement learning setting. We demonstrate our proposed method achieves the same accuracy in comparison to the standard model predictive control method using white-box models.
