Darwinian Evolution of AI skills

[Rayers12 19]

If you can implement this practically, this is the largest breakthrough in gameplay I've seen yet.

[TRexian 0]

Back from vacation... and in a private note to DM - :( .....

But, I think the tactics idea, while interesting, is not amenable to this kind of study... at least not by me. To get AI to consistently do specific things like that, in a repeatable way, it would require custom FSMs, which are currently beyond my ability.

I do think it would be interesting to compare a dataset of default AI v. a dataset of one of the AI addons.

[RobertJ 10]

This is a really interesting experiment, I don't think many games have tried 'evolving' their AI but perhaps this is a good way to get a brilliant AI.

One question, how are the new AI skills formed? do survivors keep their stats and then the new 'breeded' AI who spawn to replace the old have randomized values? One idea could be to truly 'breed' the surviving AI, making the stats an average of the two survivors since it would select to a specific 'species' of AI which could be extremely interesting. Perhaps two seperate species would form with totally different stats that survive for different reasons (low courage, high accuracy AI that don't rush in VS a highly mobile AI that can outflank its enemies and get them into a crossfire where they friendly fire each other?)

Perhaps we can have, like you suggested, Zeus AI vs ACE AI (is that possible to do?) to have a match up xD, even better you could ask for a group of players to beta test with you and we can do humans vs AI which would be a totally different selection pressure since humans are better in much different ways to AI as an AI vs AI fight (I assume) would break down into an engagement range/accuracy competiton. Maybe there should be a limit to accuracy/weapon shake to make the selection pressure go more towards tactics rather than superhuman aiming ability.

[TRexian 0]

One question, how are the new AI skills formed?

The way I set it up was to filter the survivors by those who had actually hit an enemy, that is, had accumulated 'points.' Thus, an AI who immediately fled the battle would not have his stats counted (although, in a Darwinian sense, his stats would favor survivability).

So, of those AI on both sides who successfully attacked, their scores were averaged, into 1 set of values, 1 value for each stat. Each side would then spawn AI, alternating, with each red-blue pair having the same stats, but each pair having different stats from the other pairs. The stats would be the average with a random +/-.

I haven't revisited this in a LONG time. :)

But, one issue with comparing addons is that they each may have their own way of setting the stats, along with other mechanics, like FSMs. It would be almost impossible to really compare them quantitatively, I think (although someone smarter than me may be able to come up with a way). My goal was to perhaps come up with a way for modders to have more information when tailoring the mods. Instead of, "Hey, this seems to have a good effect." they would be able to chart out how changing one skill changes the dynamic of other skills.

Anyway, glad there is continuing interest in this. :)