Sid Meier answers Slashdot

[yin26]

Well, "learning from previous mistakes" isn't a passive, mechanical process. It very much depends on what a particular human being chooses to focus on and how much talent, money and time that person (along with his team) has on hand. Many game regress in ability for these reasons.

However, Soren struck me even in the middle of the Civ 3 mess as particularly focused and capable. Also, Civ 4 development didn't suffer any trauma, so I think the talent, the money and the time were all on the game's side. Barring this, of course, Civ 4 might well have made little to no progress with its AI.

[Harrier UK]

Originally posted by vee4473


I am going to voice an opinion on this that is probably in the minority.

Sid's quote: "All in all, I don`t expect to see anything close to true human intelligence any time soon, as long as games continue to get more complex.", seems odd to me.

My first reaction was that AI should improve the more complex a game gets. I may be wrong.

But, it seems that the more variables that you throw into a decision making process, the harder it gets to make the correct decision, for most humans.

On the other hand, aren't computers much more adept at crunching large variables, odds etc... to produce the best course of action as compared to a human?

I understand that AI can't get a "feel" for what to do, but given a world where winning depends on very set rules, I would think that the AI would be able to crunch to victory all the time, barring bad luck. And the more rules and variables thrown at the human player would make the game fall more in favor of the AI.

Given the fact that the AI in previous games can't win without gross advantages, I may be wrong, but it seems odd to me.

I don't know...

If you program an AI - it can only make rational decisions - no matter how complex the variables. That is why a chess AI can beat the human.

The problem with humans is that thay sometimes make irrational decisions - even when the proper decision is obvious.

So to make a computer AI behave like a human - you need to put in a fudge factor. (i.e. attack the human tank with an AI spearman).

Then everyone complains that the AI is no good.

But in reality you may have done just that - I have. I had fortified a unit and forgot about it. 60+ turns later an enemy tank woke it up. I could not retreat, I could disband it - but why not attack the tank - nothing to lose and you hear so many stories of spearmen defeating tanks.

Get the point - a computer AI can not copy human error without being critisized as being broken.

[yin26]

Originally posted by vee4473


I am going to voice an opinion on this that is probably in the minority.

Sid's quote: "All in all, I don`t expect to see anything close to true human intelligence any time soon, as long as games continue to get more complex.", seems odd to me.

My first reaction was that AI should improve the more complex a game gets. I may be wrong.

But, it seems that the more variables that you throw into a decision making process, the harder it gets to make the correct decision, for most humans.

On the other hand, aren't computers much more adept at crunching large variables, odds etc... to produce the best course of action as compared to a human?

I understand that AI can't get a "feel" for what to do, but given a world where winning depends on very set rules, I would think that the AI would be able to crunch to victory all the time, barring bad luck. And the more rules and variables thrown at the human player would make the game fall more in favor of the AI.

Given the fact that the AI in previous games can't win without gross advantages, I may be wrong, but it seems odd to me.

I don't know...

The problem, I think, is in how the big picture gets put together. The computer can certainly crunch millions of discreet calculations together faster than any human being ever could. But this is all brute force, and the further out a consequence is, the harder it can be for a computer to reach that conclusion than it is for a human being. Of course, this all depends on how the AI is programmed to look at associations. Consider the following scene:

Two fish lay dying on the floor amid shattered glass. Next door some boys have just thrown the baseball mits on the ground and are running inside with worried looks on their faces.

A human being would almost immediately know the cause here. But a computer, unless capable of seeing a big picture, would start to piece the puzzle together literally from the pieces on the floor: counting them, categorizing them, then analyzing the fish themselves, the arrangement of the room, etc. It drowns in details.

So to the degree that the Civ 4 system allows for more paths to a conclusion, the better, in fact, the human player can put the big picture together well ahead of the computer. The classic example, of course, is Big Blue and Chess. Then, as Sid said, Chess is relatively limited in terms of possible moves and outcomes. Imagine what happens in Civ 4!

[Verenti]

To be susinct: A computer lacks creative thought.

[yin26]

Right, and the more creative a game lets you be, the more distance possible between computer and human efficiency at winning. But with Civ 4, it seems the beta testing phase was pretty strong, so hopefully a lot of those players' creative approaches were somehow captured in the AI and overall game mechanics.

[Harrier UK]

Originally posted by yin26
However, Soren struck me even in the middle of the Civ 3 mess as particularly focused and capable. Also, Civ 4 development didn't suffer any trauma, so I think the talent, the money and the time were all on the game's side. Barring this, of course, Civ 4 might well have made little to no progress with its AI.

Exactly - That is why I think they threw the towel in the ring as far as C3C was concerned.

Civ4 was already in early development and they were wasting Tavis's time ( a key member) in trying to produce a last patch to a flakey product (even the forum members could not agree on what they wanted).

So they pulled the plug on a side-project that would not earn them any revenue, and would only benefit those few people that access the forums.

IMHO the vast majority of people who buy these games are not even aware of these sites. It may be increasing but is still relatively small.

[vee4473]

points well taken guys. Just thinking out loud.

However, if the AI found fish on the floor with broken glass, wouldn't it be able to determine that that the broken glass and dead fish are a result of the opponent having the means to cause such destruction?

And most probably have known of such a possible threat beforehand and taken actions accordingly within its means?


I guess it's because the AI plans according to what its got, not according to what it "thinks" the enemy has?

The lack of foresight being the factor?

Meaning that if the AI knows the rules inside and out, which it should, it should seem like it's an advantage.

I know nothing of programming AI, so cut me some slack in these thoughts!

[yin26]

I'm no expert either. Just enjoy thinking about this stuff.

My guess is that by pure brute force, you are right. However, to actually compute all the possible variations and outcomes to at least several turns out would mean you might get to make a few turns a day.

Thus, in order to get quick turn times, the AI is constrained by varying levels of scripts...and the effictiveness of these scripts really comes back, again, to how well the human programmer can effectively make sure the computer finds that sweet spot between crunching numbers and "seeing" the obvious.

Another limit here, too, is that the human player is learning about outcomes day by day, but the computer simply reverts back to its scripts time and again, so eventually the human (if he's paying attention) develops his own scripts that reflect greater degrees of learning than what the programmer had time/talent to get in to the AI when the game shipped.

Interestingly, there is sometimes talk about "adaptive AI" in strategy games, but for all the hype, I haven't seen much come of it.

[Dactyl]

One interesting and adaptive way that the AI could be improved would be as follows. Require each user to create an identity (which he isn't locked into, of course) and, for each game that user plays, let the game track certain aspects of his play. For example, is he primarily a builder, or is he a warmonger? What units does he favor. What wonders does he tend to try for? The information that would be tracked could be determined by the deverlopers based upon their experience with the game.

At the beginning of each game, the AI would analyze the past habits of that particular player and create a profile for him, and adjust some of its strategy and tactics accordingly. Speaking as a software engineer, if this was done judiciously, it wouldn't be all that difficult to implement, and it would lead to a game that changed somewhat as you played it. For example, if you have a habit of making diplomatic agreements, then breaking them, perhaps the AI would start refusing to enter into them with you. If you played as a warmonger, you might find that the cities of the civs next to you started having more defenders. If you're big on culture, your neighbors could adopt strategies to counter that.

[Darsnan]

Originally posted by Dactyl
One interesting and adaptive way that the AI could be improved would be as follows. Require each user to create an identity (which he isn't locked into, of course) and, for each game that user plays, let the game track certain aspects of his play. For example, is he primarily a builder, or is he a warmonger? What units does he favor. What wonders does he tend to try for? The information that would be tracked could be determined by the deverlopers based upon their experience with the game.

At the beginning of each game, the AI would analyze the past habits of that particular player and create a profile for him, and adjust some of its strategy and tactics accordingly. Speaking as a software engineer, if this was done judiciously, it wouldn't be all that difficult to implement, and it would lead to a game that changed somewhat as you played it. For example, if you have a habit of making diplomatic agreements, then breaking them, perhaps the AI would start refusing to enter into them with you. If you played as a warmonger, you might find that the cities of the civs next to you started having more defenders. If you're big on culture, your neighbors could adopt strategies to counter that.

As a CMN of games and Creator this is almost exactly what I do when devloping new games! I keep a running record of what strategies players use as well as tendencies, then try to develop counter-strategies for the AI's to employ accordingly!
Anyways, I have seen this idea offered before. Surprised more people have thought of it....


D

[snoopy369]

vee ... I think it's actually not entirely true, how you are thinking.

Although most people point to the creative element as the element differing from AI to Human, I think that's not entirely true.

Humans also plan ahead more effectively than computers, in the "big picture" element. Computers in a chess game can plan ahead dozens or even hundreds of turns - but they do it by extrapolating board positions, which have a quite limited number of positions.

Civ is a FAR more complex game than chess ... and the computers we have playing it are no Deep Blue, nor are the quite impressive programmers like Soren given the time or training of Deep Blue's programmers.

I can beat an AI simply because I am more able to focus on concrete goals and big-picture strategies. It's VERY hard to ask a computer to see the big picture in a game, because ultimately you have to ask it to compare two positions and evaluate one as better, without being able to dig very deeply into the future. Who's to say if building a settler in turn 1 or building a warrior is the winning strategy? How does the computer evaluate this?

We don't know either, but instead we follow a complex and evolving strategy. AIs in Civ do this too - but they're just not as able to adjust this strategy to the big picture, to evolve it as other variables change, as we are ... thus AIs don't play complex games as well as humans.


*arbitrary comparison:
AI ability to win/draw at Tic Tac Toe: 100% (no human could ever beat an AI of any meaningful ability at TTT, and could hope only for a draw. Of course, any human could draw as well ...)

AI ability to win/draw at Checkers: 99% (a perfect AI is not going to lose at checkers essentially ever to a human.)

AI ability to win/draw at Chess: 95%

AI ability to win/draw at Civ: 20% (if you had Deep Blue playing civ against the Civ Master, on equal terms, i bet it would win less than 1/5 of its games. Heck, I bet I'd beat Deep Blue fairly frequently playing on 'no-cheating' level.)

[alva]

AFAIK, even Deep Blue was guided - while the game was going on - in a "better" direction.

[Harrier UK]

Originally posted by Darsnan

As a CMN of games and Creator this is almost exactly what I do when devloping new games! I keep a running record of what strategies players use as well as tendencies, then try to develop counter-strategies for the AI's to employ accordingly!
Anyways, I have seen this idea offered before. Surprised more people have thought of it....

D

I seem to remember reading somewhere that they did something similar to this with civ4. They played lots of multiplayer games early on and kept on adjusting the AI in areas were it struggled.

It was a Soren interview I think - but I may be wrong about who said it.