The Missing Ingredient: What’s Next for Game AI?

Alex J. Champandard on September 25, 2007

The developer discussion this week is about the missing ingredient in game AI. How can we improve the quality of AI in mainstream games? If you had to focus on one ingredient, what would it be?

Click here to join the conversation by posting a comment, and stand the chance of winning a T-shirt at the end of the month for interesting comments.

Design or Technology?

Here are some more questions to get your creative juices flowing:

  • Is technological innovation necessary to improve game AI?

  • Are there issues using existing technology in practice?

  • Does the problem lie with game design entirely?

How come there aren’t innovative games that feature intelligent/realistic/believable actors yet?

Post a comment below to join the discussion!

Discussion 8 Comments

Ian Morrison on September 25th, 2007

I guess it depends on what is considered more important: intelligence or realism/believability. While they certainly aren't mutually exclusive, an AI that always does the best possible move and uses the best possible tactics is not going to be as believable as one that engages in believable, but counter productive, behaviors. The behaviour of Halo's Grunts are a good example. Given a stimulus such as their leader dying or a plasma grenade stuck to their faces, they choose to panic instead of, say, rallying the remaining troops to a better position or rushing the player to try and take him down with them. The latter behaviours would be far more intelligent in that they increase the grunt's chances of winning the encounter (or, at least, to make sure the opponent loses as well), but those counter productive decisions go a long way towards establishing the Grunts as cowards. For creating more intelligent opponents, I think Moore's law and the trend towards multithreading is on our side. We seem to already have the tools, we just need to consciously limit them to keep processing times low. For increasing believability, though, I think that's where new technology is going to really come in handy. You've linked to a series of articles on animating AI characters, for instance. This element of AI has always been about smoke and mirrors. Speaking of which, would it be possible to have an article on here about fooling players into thinking the AI is smarter (or, as in the case of the Thief games, clueing them in on what the AI IS doing?). To me, this seems like the most important part of making believable opponents... or ones that look incredibly smart. I remember the presentation on FEARs AI. It's got a fair amount of slick AI technology (their planner is the inspiration, if not the model, for the AI in my current project), but the replica soldiers weren't doing any of the flanking, pincer attacks, or what have you that everyone thought they were... they were just finding new cover! The level designers were as important as the AI developers in that case.

diegix on September 25th, 2007

From my perspective the main thing that is limiting AI improvement in current games is the lack of time to develop it. Next-gen, at least until now, is not about getting more believable characters (I don't mean more effective) but to get bigger a prettier scenarios and characters. What is the AI function? To have the characters do stuff so that the game is a game instead of a pretty sandbox, but not too much, just enough to not be stupid. When the focus of the games switches from pretty graphics, fancy physics and huge worlds we will have the time to improve and polish AI to achieve better characters and behaviors. Most of the technology is there already: planning, learning, FSM of all kinds... The only ingredient would be the pot to mix all these ingredients and the time to let it cook with low fire. If I have to point a technology challenge, it would be terrain analysis and the use of that information to take advantage of the world. That's one of the things that human are very good at and AI characters still don't excel at.

gware on September 25th, 2007

If I had something to focus on ? I'd say "Emerging behaviors". From my point of view, what's laking in game AI nowadays is variety. Usually you can understand what's 1 unit is doing and simply paste it to the other units (of the same kind). You encounter 1 soldier, and you can bet all other soldiers will be more or less the same. It seems that most engines start to exhibit emerging behaviors, either FEAR which is doing planning in an action-space/goal-space or Halo with his "behavior tree" which is another representation of the space & travelling/transformation possibilities. Both representation tend to use "little" "fixed-rules" and try to get emerging behavior from them. Other posts on Oblivion AI system are talking about emergent behaviors (and some are focusing on how this was a problem for shipping) The problem does not necessary lies in the game design department, more like we want 2 oposites : - games shipping with NPC having well-known behaviors (so that's the game will not be buggy, can be rated 3+ , and such) - NPC being totally free. Designers want to be able to fine tune AI but they agree to say that writing down every possibilities isn't possible. The more you will set the AI free, the more you will notice emerging behaviors. And of course the more you will need resources, thus technological innovation. I want AI that will surprise me :) But still being able to win without too much efforts ;) I'm a lazy gamer ... I know ....

mihaic on September 25th, 2007

I think the missing ingredient for better advancements in Artificial Intelligence is it's more widespread adoption in other commercial applications (excepting games). Check out [URL=] this[/URL] article on my blog for more details on my opinion.

Andrew on September 26th, 2007

Wow, all good stuff to read here. I think personally, while it might not be next as in someone is doing it, I agree with Travis - characterisations is by far what I want as a gamer and AI programmer (which also part of other peoples - like improving perceptions, improving feedback). This would be excellent to make games more repayable, and if the world and design offered it, perhaps making NPC's workable beyond a 30 second life span which commonly occurs in 3d games, if they had characteristics (and the intelligence) to work out when a battle is lost, or whatnot. But varied characters, in the same enemy/ally type, would be great, especially in anything which had persistent characters through the game. For instance, I really enjoyed the different (but scripted) persona's of the 4 squad members in Republic Commando, and since they are there the entire game, and widely different personalities, a lot of enjoyment is from them talking to each other, even if the AI is the same (sadly, since they had well defined roles as a weapons expert, engineer and medic, and were not really very intelligent without direction, which was part of the fun of course). I also think characterisation can help with anything that isn't human - a lot of intelligence may be mind-numbingly different to our own, in sci-fi like ways, that characterising parts of it to act entirely contrary to a human is needed. If a general AI was built, characterisation can really helps set the aliens apart from the humans, like quite a few FPS games, and it should only get better.

Sergio on September 27th, 2007

Everyone seems to be focusing on detail AI, on specific behaviours, having more variety, better animation, etc. And this is in fact what all the studios are hard at work trying to deliver. Yet we rightly complain that the AI is lacking in many of those aspects. The reason is that trying to simulate human behaviour is an AI-complete problem, so don't expect a good procedural solution any time soon. Instead, the viable approach today is spending more time and resources on manually creating as many distinct behaviours as possible. And I believe it's not only not going to change in the foreseeable future, it's actually gotten worse with the next-gen, as more realism is required in the presentation layer, which makes building those behaviours even more challenging. If I had to bet, I'd say that the next advance in AI will come from using it in an unexpected system. Forget those boring NPCs shooting at you in all the me-too FPS games. Think of the AI driving the story of the game, or balancing it, or providing different levels to the experience. Imagine a game like GTA integrated in a world that's controlled by an RTS-like AI, with different factions warring against each other. Keep the low level simple (i.e. don't start creating hundreds of subtle variations of the same behaviour just for variety or colour) and work on delivering a different experience on the high level, give the player control over the story arcs and other macro-elements. There is a lot more room for innovation there.

Ian Morrison on September 27th, 2007

Thats a fair point. I've actually been really impressed with (and eagerly awaiting) Left 4 Dead in that regard. The "AI Director" is something that I'd mused about a little bit, but seeing it actually implemented as brilliantly as Turtle Rock has managed? That's pretty awesome. For those who aren't familiar with it, the director is in charge of monitoring the stress levels of the players and pacing the experience, as well as spawning all the zombies the players need to fight through. The resulting gameplay is (from what I've heard) unique every time, as well as highly cinematic. It's basically procedurally generated level design!

Andrew on September 28th, 2007

Ahhh, now procedural level design (or varied levels which the AI can cope with independantly, like many real time/turn based strategy games) are a whole bigger kettle of fish, although I must admit, to say it is AI related, and certainly any overarching intelligence controlling something is. Never heard of Left 4 Dead, the wikipedia article was brief so I'll check it out once released. Sounds interesting - I would say that a game which had a large enough populace involved (depending on the size of the world, presumably from a few actors (including the PC) up to millions of hypothetical ones...) could have overarching AI's which help control things which are not stimulated - being the "head" actor of a group, and therefore being able to co-ordinate the actions, would suit quite a few FPS's if each level was made replayable enough (or procedurally generated! :D )

If you'd like to add a comment or question on this page, simply log-in to the site. You can create an account from the sign-up page if necessary... It takes less than a minute!