Does the AI Learn?

[Coppermantis]

I'm playing a game with 10/10 Entrenched Homeworlder/Support Corps on a 10 planet map, and my strategy for taking planets has been first to clear out the wormhole with a lightning warhead or two, and then send in the fleet. However, recently, I've had a core world on alert for a little over an hour, and there are about 1600 ships there. I have a botnet, hive, regen and artillery golem, so if I can clear out the wormhole, I could theoretically take out the planet. (including hive wasps, I outnumber the AI).


The problem, though, is that half of all the ships on the planet are Missile Frigates! About 800 of them, double the amount of the next most common ship, bombers (400). Did the AI wise up to my warhead strategy, or is this just an unhappy coincidence?

[chemical_art]

I'm more inclined to guess the darwin effect. Repeated uses of lightning warheads cleared out non missile frigates, leaving them to gradually outnumber defenders. That is assuming you've hit that planet once already.

[Coppermantis]

Actually, now that you mention it, that's probably the reason. Now I feel dumb for not thinking of that earlier.

[keith.lamothe]

No learning going on there

[mindloss]

Nope, but as I understand it, one of the premises behind the AI is that it's supposed to trick players into thinking it's actually doing clever things. So... one point for it there.

[TechSY730]

The problem with learning AIs is that you can start playing mind games with it. Like doing one thing several times, and when they start adapting to it, you do something else, causing massive damage and making the AI look silly. Of course, the AI could try to call your bluff, but that would be complicated. And then you could try to call the AI's calling of your bluff, which leads to...

See all the issues that a learning AI faces? In fact, they are many of the same issues learning human agents face in competitive, imperfect information, not fully random games. Sadly, AIs aren't sophisticated enough to deal with these sorts of mind bending mind games, or at least not yet.

Sure, some of this could be solved by still having the AIs behavior be random, just shifting probabilities around a bit for its learning, but that is also complicated, and probably not worth it.