Defcon

[Mark]

Yesterday I attended the inaugural meeting of the Artificial Intelligence (AI) in Games Network. This is a combined effort from Imperial College London, Bradford University and Essex University to take a bunch of white-coat wearing, blue-sky thinking academics and lock them in a room with a bunch of T-shirt wearing, down-to-earth game development types and see what comes out. These collaborations are driven by the academic research councils (the chaps who pay for the research) as an attempt to try to ensure that the research outputs have some real-world applications. Now in some respects this is a very noble agenda – what is the use of research that has no practical benefit, but I believe that we do need to be careful with that line of reasoning.

There has to be a place in society for genuinely blue-sky, out of the box, crazy thinking that is unconstrained in its application. Why? Because it is not always possible to see the application of an idea when it is first conceived and developed. History is replete with examples of accidental discoveries and cross-fertilization from one field to another and I fear that if we become too focused on the specific needs of a business today, we will be neglecting the needs of tomorrow, next year or one hundred years from now.

So what does all this have to do with gaming? Well the application of AI techniques to video game design is potentially endless. As games get bigger we could use AI systems to assist the designers in creating their game worlds. We could automatically generate cities and populate the living spaces with computer styled artefacts. We could teach a system to mimic the decisions of a game designer and create basic maps that could then be tweaked to provide the perfect gaming experience. We could create Non-Player Characters (NPCs) that adapt their behaviour to match the skill level of the player or design emotional models to make the NPCs seem more realistic and human. We could work on avatars that take over for a player and mimic his style and ability. The potential is truly endless.

So why aren’t we seeing all these wonderful techniques in games now? Is it the fault of the men in T-shirts or those in white coats and what do we need to do to ease the passage of research from the lab to production-level video games? Having listened to the arguments from both sides, I’m going to take the blame and say it is industry that is providing the barriers. We constantly criticise the academics and say that their techniques would not work on real games, yet when they ask for some source code we tell them that there are “IPR” issues or that we do not have time to work with them. The very nature of research demands that most attempts will fail (or will almost certainly take longer than expected) and so we tell them that there is too much risk to let them near our precious games.

I’d like to see this change and I have a plan. Once we have finished and shipped a game, we send the source code across to the boffins with a list of super-hard problems that we failed to solve. They then use science to fix those problems and come back to us with a demonstration of their technique working. We then compare the new version of the game with the old and see if the AI has genuinely made an improvement. If it has, then we consider implementing it in the next game. It sounds simple and of course there will be details to work out, but that’s why we have an AI in Games Network.

Oh, and if anyone say’s it’ll never work – tell them the DEFCON AI is currently being turned into HAL 3000 by separate teams at both Imperial and Bradford.

[Rkiver]

Surely you mean WOPR and not HAL?

[KingAl]

Ah, very cool to hear that the Defcon AIs are being worked on... I'm guessing there are no plans for pluggable AIs in the near future given everything on your plate. AI is most certainly an exciting field, though I'm more interested in the cognitive science end - modelling of human thought as opposed to pragmatic solutions to existing problems - so I can relate to your fear of the 'blinkering' of research.

[Xocrates]

Mark wrote:

DEFCON AI is currently being turned into HAL 3000 by separate teams at both Imperial and Bradford.

Right. We will need a duel between those two AIs stat!

Then the winner will need to face the leader of the 1v1 ladder.

Finally, the winner will be placed in charge of the worlds nuclear arsenal!


Then we sit back and watch

[ynbniar]

Advanced AI? Yes please…

...sometimes even any AI will do (are you’re listening Renegade …Nort guards wandering about without a care in the world as Rogue Trooper snipers their mates is just a waste of a decent license…)

I’m interested about the Defcon AI being looked at by the boffins – if somebody can come up with an AI that is able to develop the suisilo and the STAR, organise multiple bomber swarms, and master the art of nuking pesky fighters with spare SRBM, they can go top of the class.

I think there is a lot of mileage in making characters more human…it’s like in Terminator 2 when John Conner asks Arnie if it’s possible he could stop acting like a dork all the time. Even Alex in HL2 is a bit wooden during normal play, only coming to life during cut scenes.

Then again would anyone want to play if Alex thows a hissy fit and storms off because you aren’t pulling your weight against the Antlions?

[Ace Rimmer]

Xocrates wrote:

Mark wrote: DEFCON AI is currently being turned into HAL 3000 by separate teams at both Imperial and Bradford. Right. We will need a duel between those two AIs stat! Then the winner will need to face the leader of the 1v1 ladder. Then we sit back and watch

I like the idea, if the AI is respectable. Although, the AI already uses Suisilos, it just doesn't know what to do with them.

[martin]

ynbniar wrote:

I’m interested about the Defcon AI being looked at by the boffins – if somebody can come up with an AI that is able to develop the suisilo and the STAR, organise multiple bomber swarms, and master the art of nuking pesky fighters with spare SRBM, they can go top of the class.

Of course, if the AIs are built to be effective and not play like a human at human speeds they could be quite effective at things like organising perfect nuke and bomber synchronisation - but then while that's an important part of defcon a lot is placing buildings in the right place to anticipate what other people are going to do - slightly less easy for an AI to do

[Mas Tnega]

martin wrote:

but then while that's an important part of defcon a lot is placing buildings in the right place to anticipate what other people are going to do - slightly less easy for an AI to do

Humans just use their own bastard form of statistical analysis. "That nation wins more often when it attacks with these units from this location, at this time relative to these units from that location", and so on. How hard could it be?

[Xocrates]

Much harder than that, I can assure you.

If only because humans can also assume, and react, to otherwise illogical situations. Just because answer A is usually the best, that doesn't mean it always is. The silly answer B might just work simply because they're not expecting. What you describe is basically an AI that plays the same game everytime. Making it predictable and therefore, easy to counter.

Defcon AI was damn tricky when you started to play

[martin]

it might be interesting to use a genetic algorithm to evolve building placements over time, although that would require a lot of games

[Pox]

martin wrote:

it might be interesting to use a genetic algorithm to evolve building placements over time, although that would require a lot of games

Nice idea... Start with the AI's current algorithm, then make random changes, record how successful they are and mix the good ones into the default algorithm... but yeah, that would require a lot of trials before it got anywhere sensible.

[Ace Rimmer]

But hey, don't you remember how quickly WOPR speed up the closer it game to the conclusion that the only way to win was to not play? Maybe it wouldn't take as long as you think. Just preprogram certain tatcitics for one side, like The Star, and let AI try to deal with it.

[martin]

Ace Rimmer wrote:

But hey, don't you remember how quickly WOPR speed up the closer it game to the conclusion that the only way to win was to not play? Maybe it wouldn't take as long as you think. Just preprogram certain tatcitics for one side, like The Star, and let AI try to deal with it.

well I was imagining a GA against human players, I suppose if you played GA va normal AI and then when the GA wins all the time play GA Vs GA and leave it running for a week

then we put the GA against you ace and watch it lose

[Nutter]

Pox wrote:

martin wrote: it might be interesting to use a genetic algorithm to evolve building placements over time, although that would require a lot of games Nice idea... Start with the AI's current algorithm, then make random changes, record how successful they are and mix the good ones into the default algorithm... but yeah, that would require a lot of trials before it got anywhere sensible.

yeah but what if the computer was playing against different opponents who were of different skill level and reacted differently? A tactic that would have the ai decimate a not so skilled opponent with little losses would be considered useful. Then when employed again against a more skilled opponent the outcome might be different.

And the same tactic can't be used repeatedly on the same opponent as he would know what to expect, so it would need to randomize its tactics to some degree.

... glad I don't do ai programming

[shinygerbil]

Indeed, the trouble with a genetic algorithm is how we compare success. A strategy with some excellent tactics can go horribly wrong when played up against a guy who doesn't know how to place his silos properly. So we either use a fixed strategy which is essentially like the AI DEFCON has now, another 'genetic' AI, or human players (which are fallible but much better than the AI), or a combination of all three.

Like Nutter said, I'm glad I ain't doing that.



Also, Mark said DEFCON. In ALL CAPITALS.