An optimal experimental design approach to sensor placement in continuous stochastic filtering

TL;DR

This paper develops a new gradient-based optimization framework for sensor placement in continuous-time stochastic filtering, generalizing the problem to an infinite-dimensional setting and deriving derivatives via an adjoint equation.

Contribution

It introduces a novel continuous-time OED approach using Fréchet derivatives and adjoint equations, enabling more efficient sensor placement optimization.

Findings

Derivation of the Fréchet derivative for the OED utility functional.

Formulation of an adjoint differential equation for gradient computation.

Potential for more efficient optimization methods compared to traditional discrete approaches.

Abstract

Sequential filtering and spatial inverse problems assimilate data points distributed either temporally (in the case of filtering) or spatially (in the case of spatial inverse problems). Sometimes it is possible to choose the position of these data points (which we call sensors here) in advance, with the goal of maximising the expected information gain (or a different metric of performance) from future data, and this leads to an Optimal Experimental Design (OED) problem. Here we revisit an interpretation of optimising sensor placement as an integration with respect to a general probability measure $\xi$. This generalises the problem of discrete-time sensor placement (which corresponds to the special case where the probability measure is a mixture of Diracs) to an infinite-dimensional, but mathematically more well-behaved setting. We focus on the continuous-time stochastic filtering setting, whose solution is governed by the Zakai equation. We derive an expression for the Fr\'echet derivative of a general OED utility functional, the key to which is an adjoint (backwards in time) differential equation. This paves the way for utilising new gradient-based methods for solving the corresponding optimisation problem, as a potentially more efficient alternative to (semi-)discrete optimisation methods, e.g. based on greedy insertion and deletion of sensor placements.