Quote Originally Posted by Procrustes
The problem w/ an AI - and I don't think there will be a way around this until HAL comes along - is that it can't learn, it will only ever know what was tediously programmed into it.
I can't begin to imagine how it's possible to boil down something as complex as strategy into a series of binary yes/no decisions, so I don't envy the task of the AI programmers one bit. (oops, unintentional pun there)

However, there's plenty of room on most people's HDDs in which to create additional data files so I don't see any major obstacle to a system whereby the AI can 'learn', after a fashion. For each AI faction, it records the strategic decisions they made in a campaign and the ultimate consequence of what it did in each province it owned, economic improvements added, buildings constructed, the troops it built, the moves it made with them and so on.

If it never lost a particular province throughout a campaign, it knows it can follow that same event path for that province in a subsequent campaign, 'expecting' some degree of success. Where a province ended up being lost, or the faction got totally wiped out the sequence of decisions leading to that are on record and next time, this can serve to inhibit it from going too far down that same path a second time.

Better still, a record of what the player did at any given time - especially useful where the player ultimately won - since this behaviour can be copied when you play as a different faction next time. Ultimately, you'll end up playing against 'yourself', in a way... On the other hand, the AI resolves all its battles (with other AI) by autocalc, where numbers seem to count for more than quality of troops and skill in their use. No doubt the fate of the AI trying to exactly copy what you did (including attacking large forces with small ones) will rapidly diverge from your previous success and the rest of the event tree becomes inapplicable, leaving it on its own, perhaps carving out an alternate route to ultimate failure...

With or without the ability to learn, one thing which pure applied logic can't do, which people can, is those 'what the heck' decisions about things to build or whether or not to launch an attack, say. They may have more to do with emotion (revenge for example) than reason but sometimes can still pay off. No AI would be able to anticipate such actions, nor emulate them.