Stage 2 - the real work begins

Stage 2 of our model scenario increases the existing air sector to contain a variety of aircraft and flight paths. Although we are only part way through the implementation of this stage, the increase and variety of traffic has already thrown up a number of problems related to the restrictions and ability of the search to find traffic configurations of interest.

The first big change to Stage 2 was to ensure that the emergency aircraft itself would not get resolved by RAMS.  In order to do this, we created resolution candidate rules that look for the aircraft type and decide which of the pair of aircraft will be resolved.  In a case that involves the aircraft suffering emergency pressure loss, the other aircraft is always resolved, allowing the emergency aircraft to descend to FL100 without deviation.

Stage 2 prior to resolutions rules prohibiting emergency aircraft (CPLoss) from being resolved in conflicts – here we can see CPLoss is re-routed to avoid another aircraft during its emergency descent.

As mentioned in our previous post, we have implemented a slot based mutation scheme so that traffic entering the sector on a flight path can only enter after sufficient time has passed to allow wake turbulence from any previous aircraft to dissipate.  In addition to this restriction, we have also permitted several more possible intersections with other flight paths and / or the emergency aircraft.  Bear in mind that the emergency aircraft is chosen at random from a selected high level flight path and descends to FL100.  In between these levels we have a variety of small personal jets and light passenger aircraft that can intersect with one another and the emergency plane.

This huge increase in the size and complexity of the search space has given us a number of headaches!  One is that random traffic permutations are often able to "out jump" those permutations created by the search algorithm, i.e. nudging aircraft along various flight paths by a few minutes does not always result in gaining a significantly higher risk measure than an entirely new initial configuration.  This is a feature of a very large search space and we are looking at ways to solve it, including taking a very large initial seeding of scenarios, and afterwards reducing the population to a selection.  This would in effect "bootstrap" the search in a productive direction.  However, the issue of coverage is likely to remain a vexing issue, as we have no guarantee that the mutation of a particular scenario we reject in the initial seeding would not later go on to gain a high fitness during an evolutionary run.

We will post more results as we get them.