What I'd like see AI do in SP play - Page 3

Sombre · #21 February 21st, 2007, 11:21 AM

It's not entirely a matter of memory though. I mean the game works out what happens in all battles, involving human players or not, watched or not,... so the AI players could basically calculate in advance whether it will win or not - then having done a single theoretical attack rather than an actual one, it could decide whether to commit or not. Granted this would make turns take a bit longer to process, but if you limited it to SP then I can't see that many players would complain - you'd get a dramatically less idiotic (not actually smarter) AI.

At the moment, I really, REALLY think it should be improved in a patch. I'm not expecting a huge leap in AI, but today the AI attacked a province with 45 PD with,... get this,.. a single sage. Then in the next turn it attacked with another sage and 4 militia. That's NUTS.

Gandalf Parker · #22 February 21st, 2007, 01:06 PM

Actually I culd see "test attacks" being used by the AI as being possible with the game as it is now. But it might lead to complaints about the AI cheating, using knowledge that it shouldnt have. Such as an army of units that cant usually be seen ahead of time such as glamoured units. Or commanders that have special items. It would seem hard to avoid having such things in the "thinking" process but maybe Johan could program the tests to not include items and glamoured units.

I think that the AI does do some thinking. Putting large armies of many units on the border does cause it to think twice about attacking. Part of the problem seems to be in how the AI tries to gauge the worth of its SuperCombatants. I know that there is a point formula since it comes up in balancing pretender designs, and there is plenty of disagreement by players on the results of that formula.

It appears to be an effort by the programmer to get the AI to use things that players use. Not JUST army vs army comparisons, but also equipment, SC's, and panic situations such as the seiging of a castle which seems to cause every nearby army to at least try and do damage.

Taqwus · #23 February 21st, 2007, 02:29 PM

Quote:

Sombre said:
It's not entirely a matter of memory though. I mean the game works out what happens in all battles, involving human players or not, watched or not,... so the AI players could basically calculate in advance whether it will win or not - then having done a single theoretical attack rather than an actual one, it could decide whether to commit or not.

It's not 'will win', but 'expect to win with certain troop, resource and opportunity cost against an expected opposition'. You'd have to run multiple simulations based on various scenarios (reinforcement, gem allocations, the resolution of army-push-army rules, et al).

Quote:

Granted this would make turns take a bit longer to process, but if you limited it to SP then I can't see that many players would complain - you'd get a dramatically less idiotic (not actually smarter) AI.

It can only do that calculation if it has perfect information regarding order of battle, resources, commands, and all player's orders for that turn. It doesn't get that. There's no legal in-game way for it to spy on your orders for that turn before it gives its own -- not that this has stopped other game developers like Creative Assembly, of course.

Incidentally, such a thing could also be gamed; if, for instance, I knew that a fairly strong force of mine would almost always deter the AI from attacking because it ran sims based on that army staying there, and if I also knew that a neighboring province of mine would be invaded by the same potential AI army because I'd left that province deliberately tempting, there's an obvious sucker punch to use.

Loren · #24 February 21st, 2007, 04:13 PM

I like the idea of test battles but the AI should only fight against what it knows about. Keep a record of what the AI knows of PD and the abilities of any commander the AI has faced before.

Gandalf Parker · #25 February 21st, 2007, 06:35 PM

Keeping memory isnt very likely with this version of the game. It would have to be written out each turn and then read back in.

I also figure that anything that was put in would have to be added as an optional switch. Im not sure that many solo players would want to wait while the AI tests 3 or 4 battle options in each province up to 1500 of them. And most multiplayer games dont have AIs.

Sombre · #26 February 21st, 2007, 11:49 PM

Well I don't think I'd have it test all the possible places it could attack. Instead I'd have it only test in that one place that it was going to attack, and then decide whether or not it will based on the results. This would require less than twice the length of time that it currently take to calculate. And when I say test I mean actually run the whole battle, as dom3 normally does, only as a test. Yes, this would be 'cheating'. But it isn't like the AI can use all the extra information it would get - the only thing it would effect is stopping armies of 14 people marching to certain death turn after turn. Ok, so they shouldn't really know about glamour armies before attacking - but I honestly don't see why anyone would need the glamour advantage against the AI. I mean right now in order to beat it all you have to do is let it kill itself off against your PD while you build up forces. That removes a lot more strategy than losing the ability to 'trick' it with glamourous defence forces.

Unlike most 'exploits' that you can use versus the AI this lemming weakness that they have isn't something you can choose not to exploit if you prefer a more challenging game. I mean you could choose not to use PD at all against the AI, but that's pretty extreme and you'd still be guarding your provinces and they'd still be suiciding into them. So what can you do, refuse to defend your borders at all?

If it used this form of test before attacking you could basically forget you knew about it and not use it against the AI - it is still deterred by strong PD and armies, but eventually it's going to attack with a much more built up force. Plus the AI would still make mistakes - since the outcome of battles is somewhat based on a series of random numbers there's room for it to get the test wrong. You could also tell it to only perform the test 80% of the time or something.