So How has the AI improved ??

[Didz]

What comes across from your post is that a game that's "on the line" is decided by "the hand of god" in TW. It isn't.

Nope! its determined by the combat sub-routines random variable generator, God has nothing to do with it.

The point I was trying to make, and have been since the start of this debate, is quite a simple one. Namely that if you present the AI with a totally balanced scenario where neither it nor its opponent has any advantages then any action the AI takes cannot be a test of its performance because any risk assessment it performed would be unable to establish a preferred reaction.

I realise that I have failed miserably to get this point across but I shall have one more attempt before I give up.

Lets look at an extreme example:

Two opposing but identical spear units facing each other across a totally flat peice of terrain. No wind, no sun in eithers eyes, absolutely identical conditions and stats. The only difference between each is the colour of their uniforms. The AI assesses the situation and comes up totally blank. There are no advantages to be gained by moving, there are no disadvantages either. If it attempts to move to a flank the enemy will change frontage. Worse case scenario is that if it moves it loses unit cohension therefore it stands and does nothing. Therefore, its default 'do something' sub-routine kicks in and forces it to perform a random action. It advances, the result could go either way, but the result was not initiated by the AI but by the default 'do something' routine which overrode the AI to make it attack.

Lets look at two other examples:

CHESS PROGRAMS: Chess programmes probably have the most sophisticated AI routines in the gaming industry, some can even beat grand masters. But no chess programme AI routine can generate an opening move, simply because at the start of a chess game there are no advantages or disadvantages to compare. The board is totally balanced and so if given the problem an AI routine will go into introverted analysis for ever in the hope of finding one. To avoid this Chess programmes are given a database of standard opening moves, sometimes five or more moves in a sequence, which the program will select from at random as a 'do something' strategy. So, in the first few moves of most Chess games against a computer you are not actually playing the AI at all by a random move sequence hard wired into it.

and finally an example from an AI used in real warfare:

Onboard ship to ship defence systems are heavily control by computerised auto threat management and response computers, with highly sophisticated AI systems. However, even they cannot make decisions based upon a totally balanced scenario. This situation was never considered likely to happen in real warfare and so was initially overlooked, until a ship found itself under attack from two identical missiles fired from exactly the same distance but from different directions. The AI routine went into 'assessment mode' looking for any factor which would help it decide which missile to deal with first and continued to monitor the alternative threats but do nothing until both missiles hit the ship. Now all such battle systems have built in 'do something' sub-routines which will kick in and override the AI to force the ships system to attack one or other of the threats at random.

That was really the only point I was trying to make, and I could not see how one would judge the performance of an AI in a situation where effectively it isn't at the helm but has been overridden.

It not an important point, but if one is then going to judge the AI on what happens its worth noting that in fact it probably isn't responsible for the end result. Its actually the 'do something' sub-routine that your testing.

However everyone seems to expect that the AI will beat them at their own game of exploiting this or that: fair enough. You might be waiting more than you think though from my perspective.

Certainly the best AI would be one that monitored both its own and its opponents performance and adjusted its own strategies to compensate.

In other words an AI which learns from its own mistakes and adjusts its own playing strategies to expliot not just the immediate situation it is in, but the long term behaviour of its opponents.

There have been some attempts at this sort of AI before but so far its yet to become common in gaming.