Generalising Cost-Optimal Particle Filtering

TL;DR

This paper introduces a generalized framework for cost-optimal sensor scheduling in dynamic systems, balancing resource use and estimation accuracy through active observation decisions.

Contribution

It formalizes a flexible model for sensor scheduling that minimizes costs while maximizing state estimation quality, extending prior methods.

Findings

Formalization of a general sensor scheduling problem

Solution to two example scheduling problems

Discussion of future research directions

Abstract

We present an instance of the optimal sensor scheduling problem with the additional relaxation that our observer makes active choices whether or not to observe and how to observe. We mask the nodes in a directed acyclic graph of the model that are observable, effectively optimising whether or not an observation should be made at each time step. The reason for this is simple: it is prudent to seek to reduce sensor costs, since resources (e.g. hardware, personnel and time) are finite. Consequently, rather than treating our plant as if it had infinite sensing resources, we seek to jointly maximise the utility of each perception. This reduces resource expenditure by explicitly minimising an observation-associated cost (e.g. battery use) while also facilitating the potential to yield better state estimates by virtue of being able to use more perceptions in noisy or unpredictable regions of state-space (e.g. a busy traffic junction). We present a general formalisation and notation of this problem, capable of encompassing much of the prior art. To illustrate our formulation, we pose and solve two example problems in this domain. Finally we suggest active areas of research to improve and further generalise this approach.