A component of the safety net ensuring the safety of air traffic is Short Term Conflict Alert (STCA) systems, which use ground-based radar to monitor aircraft and alert air-traffic controllers to potential airspace conflicts, who can then direct pilots to take suitable evasive routes. The National Air Traffic Service's (NATS) STCA system covering the busy London terminal airspace handles over 3000 aircraft per day and its importance is highlighted by the fact that it is thought that one of the factors contributing to the midair collision over the border between Germany and Switzerland in July 2002 was that parts of the STCA system in the relevant Swiss control station were switched off for maintenance. It is crucial to raise alerts for the 5-15 daily over London genuine potential airspace infractions, true positives, but important to minimise the number of false positives so as to avoid crying wolf too often to air traffic controllers who would tend to lose confidence in the system.

The STCA program is a vastly complicated predictive model with more than 1500 adjustable parameters, such as `vertical closing rate threshold', which are currently manually adjusted by skilled NATS staff using a database of about 170000 historical and recent encounters. This laborious tuning must be carried out in response to changes in volume of air traffic, local traffic operational procedures and changes in the regulatory environment.

As part of an EPSRC programme on Critical Systems and Data-Driven Technology, we have used evolutionary algorithms for multi-objective optimisation to locate the Pareto front---the curve that describes the optimal trade-off between the true positive and false positive rates. Multi-objective optimisation acknowledges that when attempting to maximise or minimise two objectives (for example, price and performance, or here true positive rate and false positive rate), there will often be optimal solutions that are not wholly better or worse than each other; for example, solution A may have better false positive rate than B, but worse true positive rate and vice versa. Recent advances in evolutionary algorithms permit the efficient location of the Pareto front of solutions which are not wholly worse than any other solutions. The picture shows solutions marked as red dots on the Pareto front describing the optimal trade-off between genuine and nuisance alerts for the London airspace.

The warning time given to air-traffic controllers of an impending serious encounter is a third objective that may be optimised simultaneously with the genuine and nuisance alert rates. The resulting Pareto surface describing the trade-off between the three objectives is visualised in the picture below by colouring solutions on the two-dimensional front according to the warning time.