Variable interactions and exploring parameter space in an expensive optimisation problem: Optimising Short Term Conflict Alert

Abstract

Short Term Conflict Alert (STCA) systems provide warnings to air traffic controllers if aircraft are in danger of becoming too close. They are complex software programs, with many inter-dependent parameters that must be adjusted to achieve the best trade-off between wanted and nuisance alerts. We describe a multi-archive evolutionary algorithm for optimising regional parameter subsets in parallel, reducing the number of evaluations required to generate an estimated Pareto optimal Receiver Operating Characteristic (ROC), showing that it provides superior results to traditional single-archived algorithms. A method of `aggressive' optimisation, designed to explore unknown parameter ranges in a `safe' manner, is shown to yield more extensive and better converged estimated Pareto fronts.