AI on Cell

[pc999]

A interesting interview

The local memory of the SPE could allow a core to work on say, a constantly running learning task, not disturbing the rest of the system.We can imagine in a RTS that a SPE will continuously monitor the player perceived actions (with no cheating) and will emit recommendations to the rest of the AI (running on another SPE) such as:

* "The player is most probably building an air force ! Be prepared !"
* "The player is most probably behind us in its resources discovery ! Exploit that !"
* "Considering the player habits, this wave of tanks will most probably be followed soon by a wave of planes ! Be prepared !"

I remember that some time ago MS has been searching for programers to do predictions and such, It should be a very interesting gen in AI terms...

http://www.psinext.com/index.php?ca...mp;PHPSESSID=b65d92f019a593f80cb2485b0fa2d1fe

 

[3roxor]

* "The player is most probably building an air force ! Be prepared !"
* "The player is most probably behind us in its resources discovery ! Exploit that !"
* "Considering the player habits, this wave of tanks will most probably be followed soon by a wave of planes ! Be prepared !"

Isn't this how AI is done now :? If you play an RTS game against the computer you will almost never succeed using the same strategy twice in a game.

 

[Titanio]

If you play an RTS game against the computer you will almost never succeed using the same strategy twice in a game.

Hehe, if only that were true. In my experience, the opposite is the case (well I guess it varies from game to game, but with one of the last I picked up - C&C Generals - the computer opponents were very predictable).

 

[pc999]

[Isn't this how AI is done now :?

Actually I think that today is more :

There is a wave of planes ---> predefined reaction.

 

[Guden Oden]

Isn't this how AI is done now :? If you play an RTS game against the computer you will almost never succeed using the same strategy twice in a game.

Hm... Pretty much the only computer game that has more than even half-decent AI without resorting to blatant cheating is chess.

Typically, the way to make RTS AIs hard is to modify the rules for the computer; reduce build times, increase resource extraction rates etc.

The big trick will be making AIs that play realistically and challenging the player without becoming impossible. There must be openings for the player; the computer already has the advantage by total battlefield awareness and supreme micromanagement abilities. It can run battles on two, three, five or more fronts without breaking a sweat, but have a player handle two separate battles, PLUS manage build queues in factories AND base construction at the same time... Very very taxing and stressful.

I would like to see computer AI that is set under the same restrictions as a human, IE, can only view a screen of map at a time (plus whatever shows up on minimap through fog of war), plus having a reaction time. Computer players typically react INSTANTLY, this is pretty much cheating actually...

 

[3roxor]

That is also what I meant..the computer basically just looking at what you have build and then (out of your view) cheats just enough counter units(possible numbers offcourse) just to make it hard on you. Offcourse this isn't real AI at all.