PSO for completion time improvement?

[Lotharsson]

Has anyone had success using the particle swarm optimizer for improving completion time (e.g. on Speedway or The Gap) ? I can't recall any success - from (possibly faulty) memory they tend to end up getting stuck at a significantly worse completion time.

[TechSY730]

Remember, when pulling the best car out of a PSO, make sure to use the Best Car button. As you imply, the PSO does not guarantee that each generation will contain something as good or even close to the best of the previous generations. The best car button will export the best car found over all generations, not just the current one. And with the evaluation logic, any ties in score should be broken by which one has the lower completion time, so it should return the fastest. (If it isn't then there is a bug)

It can be frustrating to unable to tell whether the PSO is actually finding improvements to time at a glance; a request to add a time to completion part to the history graph has already been made.

[Lotharsson]

Yes, I use Best Car to pull the best one out of PSO. So far I don't think it's ever been any faster than my input seed - and the population tends to converge almost entirely to cars running at a particular slower speed - or failing to get any significant distance at all. This state of affairs seems to persist over hundreds of generations, at least in the instances that I recall.

[TechSY730]

PSO with a mutation rate of 0 does tend to converge to a local maximum after a couple dozen of generations. Interestingly, it does respect the mutation rate. Try bumping it up a bit, and there is a chance that a better car could be found pushing the PSO away from the local maximum it is currently obsessed with.