Nonlinear Optimization Filters for Stochastic Time-Varying Convex Optimization

TL;DR

This paper introduces nonlinear filtering-based algorithms for stochastic, time-varying convex optimization problems, demonstrating their effectiveness through real-world ride-hailing data, especially with accurate predictions.

Contribution

It develops novel online algorithms using nonlinear filter design principles, including extended Kalman filters and bilinear matrix inequalities, for tracking optimizers in stochastic time-varying convex optimization.

Findings

Algorithms perform well with accurate predictions.

Numerical results show effectiveness on real ride-hailing data.

Certificates provided via LPV analysis and sum-of-squares relaxations.

Abstract

We look at a stochastic time-varying optimization problem and we formulate online algorithms to find and track its optimizers in expectation. The algorithms are derived from the intuition that standard prediction and correction steps can be seen as a nonlinear dynamical system and a measurement equation, respectively, yielding the notion of nonlinear filter design. The optimization algorithms are then based on an extended Kalman filter in the unconstrained case, and on a bilinear matrix inequality condition in the constrained case. Some special cases and variations are discussed, notably the case of parametric filters, yielding certificates based on LPV analysis and, if one wishes, matrix sum-of-squares relaxations. Supporting numerical results are presented from real data sets in ride-hailing scenarios. The results are encouraging, especially when predictions are accurate, a case which is often encountered in practice when historical data is abundant.