Actually I agree with Raven. 

Me too. Of all aspects of Elemental, the AI is the part I'm least worried about. Not because it isn't important - it's one of the most important - but because Stardock/frogboy have proven that they are among the best when it comes to TBS AIs. And perhaps even more incredibly, they have a track record of continually improving their AI in huge ways.

This.  Moreso than trying to make the perfect AI I would love to see lots of variety.  Some that are super agressive, some that build super cities, some that focus on spells and of course some that are more balanced.  The one thing that kills AI more than anything is predictability.  Not sure if you guys have the resources to make many personalties but that's what I would like to see.

Hehe maybe. The reason I say it will be better in a year is because of what I've seen Stardock do with Gal Civ 2. They constantly upgraded the AI. By the time the last expansion came out the AI went through a drastic change from launch.

As I want to compete in the metaverse and post topscores, A.I modding is not interesting to me.

I hope that Stardock makes it so people can't edit the A.I and then abuse it in the metaverse.

A chess board has 64 tiles.

Its "AI" is a simple matter of calculating the value of different pieces and holding different positions.

Game AI does not remotely resemble chess.  My AI doesn't have the luxury of looking through thousands of tiles. It would take hours for each move even on a modern PC.  I have to implement strategies based on my judgment of what is and isn't a good tactic to take.

The main reason game AI tends to be rotten is because most developers aren't necessarily very good at their game.  

Should an AI spend its resources on recruiting an NPC? Or building a new unit? Or upgrading its city? Or equipping his sovereign at the shop? Or giving resources to a friendly AI that is at war?  There is no "mind boggling about of calculations" that can determine what is the "best" thing to do. 

And considering that when I was a kid we used fountain pens instead of ballpoint, slide rules instead of calculators, I first watched star trek in B&W not color let alone HD (or soon-to-be 3D), there were only 3 networks, electronics used tubes (all of them, not just the top-line), you could get a handful of candy for 1 penny, etc. etc. etc. -- I wouldn't bet on your "lifetime" statement, as I've still got a good 30 years to go (if my relatives are any indication).

Or maybe I just have more faith in programmers...

The main reason game AI tends to be rotten is because most developers aren't necessarily very good at their game.

It's possible to create systems that learn without necessarily understanding what they've learned, but such systems can produce very bizarre results if you don't give them well-structured training.  Neural networks are one fairly well-known example of this type of system.  A neural network couldn't execute the logic for an entire AI player (the set of inputs and outputs is too large to quantify in any reasonable way), but they could be interesting for simpler decisions, like deciding on and prioritizing high-level goals based on some representation of current known world 'state'.

The designer would still need to make a lot of choices as to how to quantify the state such as 'threat level from player A', 'my economic performance', 'available materials', etc.  The designer would also need to specify the set of available goals, such as 'develop industry', 'build up army', 'build up tech', etc.  The designer would also need to design how the different assigned weights to the various goals results in actions being dispatched on each turn.

The main thing the designer doesn't specify directly is any meaningful logic that updates the goal weights based on the state inputs, as this is determined by the neural network, which has been trained to roughly correlate particular sets of inputs with particular outputs.  The designer still introduces bias in the way that input/output training sets are chosen, but you can have human players (not necessarily the designer) play the game to generate the training sets.  Good players playing well become the primary training sets that initially tune the network.  Done correctly, this can have interesting implications, such as the ability to adapt a player's own strategies (tit-for-tat type of play) and also the ability to train different 'styles' of play by observing different styles of human player.

The main problem with these systems is that they can be quite unstable and produce very strange results for inputs that are outside the scope of the original training set or that lie in certain 'special' places in the input set.  I'm curious if they've ever been used in games.

Who says that's a flaw?  I know if someone killed one of my offspring (interesting way to think of them), I'd hunt them down to the ends of the Earth or die in the process.

But yeah, being able to make an ai have qualities like that among others would be great. 

One of the best parts of LAN (or voice) MP with old friends is hearing the lamentations of their women... errrr... hearing the wailing and gnashing of teeth when a battle goes against them.  AI that teeth-gnashed and   would be fun.