Real time AI = bad, Turn-based AI = Good, why?

[dstebbins]

Why is it that, in video games, AI that is real-time is easy to defeat, yet AI in turn-based games is virtually impossible to beat?

 

I mean, I've played racing games against AI-controlled cars, and come in first place, yet, when I go online and race real people, I end up coming in dead last.

 

Yet, I find that I have a LOT better luck kicking a real person's ass in Chess than I do a computer.

 

What makes AI take such a 180 in difficulty?

[ecoli]

speak for yourself... I find (for example) the AI in Starcraft very difficult but in CIVIII easy. Mostly because, I think, the AI can move faster than I can. The mouse really slows me down, and I like to think. I'll stick to turn based-games.

[CharonY]

In all games it depends on the mechanism. Being RT and round based is but one parameter. With limited option and easy evaluable outcomes good AIs are likely to make less errors. The Civ games are so open, however that often the generalized, often non-adaptive strategies will lose after one gets the hang of it. In RTS reflexes will be an issue and creating a game where one has to be everywhere at one will put humans at a disadvantage (though games are of course designed to be controllable to begin with). I have heard good things about the Galactic civilization series (I think it is called that) and of course highly sophisticated chess routines have been shown to best the human best chess players.

 

Races are yet another issue. It is possible to program the computer drivers to be perfect (i.e. running perfect laps using the best route etc.). Only it wouldn't be any fun to play. And that is the final bit, computer games are intended to be challenging, but also enjoyable. Even in games where it would be easy to create perfect adversaries, it would not be fun to do so.

[Mr Skeptic]

Another aspect is that for realtime games the AI matters very little. In various of the old arcade type games, the "AI" of every creature consisted of a very simple repeated script, yet it was still challenging because reflexes played a large role. Basically the AI does not necessarily need any intelligence in some types of games. Also the purpose of the AI is to provide fun and challenge to the game player, not to be skillful and intelligent. Keep in mind that many AI players also cheat, such as by having a resource bonus.

 

Chess is rather a different story; it is a very closed game which gives computers a huge advantage. A computer can beat the best human chess player, and a computer has "solved" checkers (is unbeatable in theory). But even among board games, the more open-ended games such as Go and Backgammon are harder for computers. And for large open-ended games like the Civ style games, I can beat their turn-based AI even when they have an impressive resource bonus and smartest AI, and that is especially easy for me on larger maps.

 

I prefer the turn-based games, because it forces the developers to focus on wits rather than on reaction times.

[Sisyphus]

Keep in mind that many AI players also cheat, such as by having a resource bonus.

 

 

Indeed, I've played some games where the AI doesn't get any smarter (and in fact remains quite stupid) on the more difficult settings, it just cheats more. Civilization Revolutions is a particularly egregious example of this, where on the most difficult setting dozens of enemy armies will just materialize outside your cities every turn, and the combat result "dice" are loaded heavily in their favor, yet they remain incapable of strategy that would be obvious to any human. It's as difficult as playing an actually skilled human opponent, but completely different and ultimately much less interesting. [/nerd rant]

[swansont]

I agree — in CIV the AI cheats, and this is more egregious at the harder levels; the AI knows where resources are going to appear, and which cities are poorly defended.

 

I think it's a matter of whether the AI is written so it can leverage the advantages of the computer. In Starcraft and Warcraft, the computer can keep track of all of its units and prioritize actions, and carry them all out in real time, i.e. it has very strong logistics. Back when I played those games, I had trouble once I reached the point where there were several dozen units and simultaneous battles were taking place, along with resource gathering and building improvements. The computer would start a new build just as soon as the old one ended, but I would be distracted by something and slow to react. And in Warcraft I could never get the archers to stay back and use their strength (orders? Shmorders!) — they'd start fighting hand-to-hand, and get slaughtered. There's no way to individually control many units in real time.

 

Some AIs have really poor strategy and tactics, though, and it's often hard to make up for that with superior logistics. If the computer can't accurately anticipate what you might do, it can't prioritize its own actions.

[Bignose]

But even among board games, the more open-ended games such as Go and Backgammon are harder for computers.

 

I have no idea where it is today, but about 7 or 8 years ago when I played backgammon far more regularly, the best computer programs could beat all but the best 15 or humans in the world, over the long term. "Over the long term" is needed because the random element of the game means that in almost any position, both sides have a finite chance of winning. Computers were able for every move to be able to choose the move that always gave their side the greatest chances of winning. Again, several years ago, there were really only about a dozen or so people who knew the game well enough to beat the best computer programs. I have to imagine that it is almost impossible today, considering the consistent increase in computing power.

[Mr Skeptic]

I have no idea where it is today, but about 7 or 8 years ago when I played backgammon far more regularly, the best computer programs could beat all but the best 15 or humans in the world, over the long term.

 

Hm, you're right. I think I confused "hard to solve" with "hard to play well". The dice make the "simple" minimax approach infeasible, which is probably what made the game interesting to computer scientists. Apparently modern programs use a neural net for backgammon.

 

Go however is still hard for computers, especially the bigger board: http://en.wikipedia.org/wiki/Computer_Go

Still, they are making very good progress too, and I guess it won't be long before they beat the best humans.

[dragonstar57]

but how long until they can command real armies of their own???