Systematic Generalization for Predictive Control in Multivariate Time Series

Prior work has focused on evaluating the ability of neural networks to reason about novel combinations from known components, an intrinsic property of human cognition. In this work, we aim to study systematic generalization in predicting future state trajectories of a dynamical system, conditioned on past states' trajectory (dependent variables), past and future actions (control variables). In our context, systematic generalization implies that a good model should perform well on all new combinations of future actions after being trained on all of them, but only on a limited set of their combinations. For models to generalize out-of-distribution to unseen action combinations, they should reason about the states and their dependency relation with the applied actions. We conduct a rigorous study of useful inductive biases that learn to predict the trajectories up to large horizons well, and capture true dependency relations between the states and the controls through our synthetic setup, and simulated data from electric motors.