I agree with the three points and would like to add another.

Design the AI to win. Not to come in second place.

First and foremost, the AI needs to manage its economy efficiently - this includes city development, resource use, research, unit design and production, and a dozen other things - without needing to cheat horribly. If it can also be taught basic military strategy, in other words how/where to use the units it has most effectively, that would be amazing - but first it needs to build an empire capable of competing militarily, economy comes before all else.

It's fine to say all this, but to codify it in an AI routine is entirely another matter. A lot of this stuff involves qualitative analysis which is extremely difficult for an AI to do.

joasoze reminded me of the Art of War. In it, I seem to recall a line that goes:

Which gave me a further idea on handling wartime with the AI. Once the AI decides to go to war against a player, it should consider (perhaps randomly) some "recipes" for conducting war. Selecting one of them, it should use it during the war.

One such "recipe," would be to instruct the AI to create two forces: a small one and a large one. Send the small force toward one city, wait until the player has moved troops toward it and begun to mobilize their troops to deal with the "threat," then move the large force on a different (more important) city. Effectively this would be an example of the AI deciding to "bait" the player and "draw agro" from the PC. : )

This isn't something the AI can "figure out" on it's own. It needs a step-by-step deterministic guide to do it, with rules with what to do in certain cases. "Little force hasn't been attacked after ten turns - start pillaging countryside." "Big force just got scouted by the player and was within his fog of war, abandon plan and attack normally." etc. That's what I'm calling "recipes" in this game: nuggets of good ideas that can carried out quantitatively.

A few to start:

-Exploit your advantages, avoid your weaknesses -- using knowledge gained by non-cheating methods, construct your forces to counter your foe's strengths and avoid her/his weaknesses.

-Hit them where they're weakest, where they don't anticipate (not just the nearest).

-Avoid their strengths (withdraw when a strong enemy force approaches, leading them around so as to neutralize the force)

-Don't attack until you have sufficient forces to accomplish a major goal (take a city or a few cities -- enough to seriously crimp your foe).  To do otherwise just tips your hand and gets your foe to mobilize, making subsequent attacks less likely to succeed.

-Identify your foe's strategic resources/production/etc. and prioritize taking it out.   Make this a goal (see above)

-Be aggressive, to keep your foe off-balance.  Try to determine where/when the attacks occurs.  Keep the initiative.

-Information is vital -- scout, spy, recon, etc.  (hopefully the AI won't cheat)

-learn from previous battles, and expect your opponent to do also.

-vary attacks so your foe doesn't fine-tune their defenses/attacks, and to keep your foe off-balance

EDITED IN:

-choose units to sacrifice (soak-off in grognard terms).  In Civ4 catapults can be routinely sacrificed to weaken a foe, allowing other units to survive and gain experience.

-Don't have the AI always attack with the unit that has the best odds.  Another Civ4 example -- the game will choose to have a highly-promoted but weakened player's unit defend when there are much less valuable units available.  Example from a game a few weeks ago -- I took a city and garrisoned with a dozen units, one unit was very highly-promoted and damaged, the rest unpromoted.  One enemy unit attacked.  Even tho the city couldn't have fallen the game had the best odds unit -- the damaged highly-promoted unit -- defend in that meaningless defense when there were many other relatively valueless units available.  The good unit died. 

Units should have a 'preservation value' assigned, allowed to over-ride basic attack/defense odds, dependent upon the seriousness of the battle (defense of capital city and Sov -- highly important, defense of small stack or in an attack where no goal (ie - capture city, destroy resource, etc.) is achievable then highly unimportant.

-attack order is important -- AoE units first, less-valuable units first, etc.  Again Civ4 -- how often will a stack have swordsmen attack then catapults attack, instead of the reverse?

Not being a programmer or game designer, I may be way off on what's possible.  That won't stop me from offering ideas, in the hope of my learning.

Regarding algorithims/etc., an AI that learns, game design for combat, etc., what's wrong with this sort of idea:

Suppose we're starting with a simple, typical battle -- city siege.  And we're making simple assumptions (troops essentially equal on both sides, no special abilities to start, etc.) and using 'average' capabilities (wooden walls, moderate unit strength/armor/hits/etc.). 

Let's also make simple definitions -- each side's strength is the sum of its unit's 'hit points', the fight lasts until one side is eliminated (the loser), if more than half the winner's strength remains it's a significant victory (and a significant loss for the other side), etc.  And let's assume the goal for the attacker is to have 60% of it's strength remaining.

If the attacker loses significantly, add 50% to attacker's starting strength for next time.  If the attacker loses normally add 25%.  If the attacker wins but has less than 60% strength remaining add 10%.  If attacker has more than 60% remaining subtract 10%.

Start with defender having 100 strength.  Make an assumption on what strength attacking force is needed to start then run 1000 attacks, with the AI 'learning' from each to modify starting attacker strength for the next attack.  You might end up with a pretty good estimate for that scenario. 

Toss out the first 500 attacks (assume it's still 'calibrating') then look at the last 500.  Does the determined attack strength jump around a lot?  If so there may be a problem with randomization of combat results, or some similar problem.  Run an ANOVA (or whatever they're doing these days).  Maybe the results are too consistent.  If so, add more variability.  etc. etc. etc.

Then repeat the above with defender strength of 200, 400, 800, etc., and 50, 12.  This'll determine the effect of defender streng and allow adding weights to adjust accordingly.

Then add in more specific capabilities, and repeat.  You may be able to make correlations, such as -- adding in 50% of the defender's strength as ranged equals doubling standard defender strength.  This will help create algorithims for army composition.  Similar comparisons can be made for other abilities, such as fire spells vs units resistant to fire, etc.

While the above is from the point of view of the attacker, it can be used by the AI on defense (by holding attacking strength unchanged and modifying defender strength).

Having an AI that is able to 'learn', and being able to run thousands of simulations very quickly/easily, for a number of different situations -- won't that allow some good guesses for constructing algorithims to make a decent AI?

That makes me think of the movie Wargames. Let the AI play itself at tic-tac-toe until it realizes that no one will win in the long run, so it should try diplomacy instead. Or maybe have it run battle simulations a few times to give it a value for "what are the odds I'll win?". Then the AI personality can determine if it fights or runs. Some will run at low % chance, while others will have "more of a taste for risk".

The three things you posted in the first post happen at empire level: know when not to engage, how and when to expand, and which objectives to hit. Those are really hard to code decisions.

About combat, that's pretty hard too. Auto-resolve most times gives horribly wrong results, so an AI that uses it won't be very good actually. Really, if you have coded a professional game AI, you can't be saying these aren't complex problems...

I'm not saying it has to be flawless, I'm saying it is a very hard problem to solve well.

And an AI using that calculation will only be good if the auto-resolve gives good results, but all auto-resolves I remember give results that aren't very similar to the real result if the player is involved.

VicenteC has a legitimate concern.  If AI is just hard-coded then it's very important the devs create a challenging AI right out-of-the-box.  Flashy graphics attract folks to try a game, great lore will help keep folks playing, but lousy AI will consign a strategy game to the bargain bin pretty quickly, even if it has good MP and dedicated modders.

Except when combat falls on the gray zone where AI thinks it is going to win with auto-resolve (or inflict important damage) and it gets massacred in normal tactical combat (which it is not a strange scenario in games with tactical combat). For extremes, auto-resolve works, but for the cases in the middle, simply assessing troop strengths = inaccurate results = you have an AI that takes bad decisions.

Thanks!

Human participation will have some effect that may throw off C vs C results (depending on the relative skills of AI&players).   Won't your "...run the autoresolve hundreds or thousands of times per round depending on the algorithm used." be just C vs C, and if so, how will you account for human participation (which is the only point of his I'm addressing) without some 'fudge factor'?  Without some 'fudge factor' I think his concern is valid.  With it (or its equivalent) I agree with you.

Hyperbole aside, I don't think you two (errr...  us three) are in that much disagreement and it's more a matter of focusing on slightly different aspects of the same basic point.