Multiple-objective sensor management and optimisation

Abstract

One of the key challenges associated with exploiting modern Autonomous Vehicle technology for military surveillance tasks is the development of Sensor Management strategies which maximise the performance of the on-board Data-Fusion systems. The focus of this thesis is the development of Sensor Management algorithms which aim to optimise target tracking processes. Three principal theoretical and analytical contributions are presented which are related to the manner in which such problems are formulated and subsequently solved.

Firstly, the trade-offs between optimising target tracking and other system-level objectives relating to expected operating lifetime are explored in an autonomous ground sensor scenario. This is achieved by modelling the observer trajectory control design as a probabilistic, information-theoretic, multiple-objective optimisation problem. This novel approach explores the relationships between the changes in sensor-target geometry that are induced by tracking performance measures and those relating to power consumption. This culminates in a novel observer trajectory control algorithm based on
the minimax approach.

The second contribution is an analysis of the propagation of error through a limited-lookahead sensor control feedback loop. In the last decade, it has been shown that the use of such non-myopic (multiple-step) planning strategies can lead to superior performance in many Sensor Management scenarios. However, relatively little is known about the performance of strategies which use different horizon lengths. It is shown that, in the general case, planning performance is a function of the length of the horizon over which the optimisation is performed. While increasing the horizon maximises the chances of achieving global optimality, by revealing information about the substructure of the decision space, it also increases the impact of any prediction error, approximations, or unforeseen risk present within the scenario. These competing mechanisms are demonstrated using an example tracking problem. This provides the motivation for a novel sensor control methodology that employs an adaptive length optimisation horizon. A route to selecting the optimal horizon size is proposed, based on a new non-myopic risk equilibrium which identifies the point where the two competing mechanisms are balanced.

The third area of contribution concerns the development of a number of novel optimisation algorithms aimed at solving the resulting sequential decision making problems. These problems are typically solved using stochastic search methods such as Genetic Algorithms or Simulated Annealing. The techniques presented in this thesis are extensions of the recently proposed Repeated Weighted Boosting Search algorithm. In its original form, it is only applicable to continuous, single-objective, optimisation problems. The extensions facilitate application to mixed search spaces and Pareto multiple-objective problems. The resulting algorithms have performance comparable with Genetic Algorithm variants, and offer a number of advantages such as ease of implementation and limited tuning requirements.

More information

Identifiers

ORCID for Neil White: orcid.org/0000-0003-1532-6452

Catalogue record

Export record