Neural Control Systems

TL;DR

This paper introduces a control-theoretical approach to function learning using reproducing-kernel Hilbert spaces, deriving gradient-based optimization methods that are effective on toy and real-world problems.

Contribution

It presents a novel control-theoretic framework for function approximation, including gradient derivation and optimization techniques aligned with Pontryagin's maximum principle.

Findings

Gradient-based control optimization effectively learns functions from data.

Methods work on toy and real-world high-dimensional problems.

Approaches are versatile across different learning tasks.

Abstract

We propose a function-learning methodology with a control-theoretical foundation. We parametrise the approximating function as the solution to a control system on a reproducing-kernel Hilbert space, and propose several methods to find the set of controls which bring the initial function as close as possible to the target function. At first, we derive the expression for the gradient of the cost function with respect to the controls that parametrise the difference equations. This allows us to find the optimal controls by means of gradient descent. In addition, we show how to compute derivatives of the approximating functions with respect to the controls and describe two optimisation methods relying on linear approximations of the approximating functions. We show how the assumptions we make lead to results which are coherent with Pontryagin's maximum principle. We test the optimisation methods on two toy examples and on two higher-dimensional real-world problems, showing that the approaches succeed in learning from real data and are versatile enough to tackle learning tasks of different nature.