This is the controversial one. Emergent behaviours are non-scripted, un-planned activities that intelligent game characters learn through their interactions with the game world. Even the programmer won’t know what they’re going to do. There have been some amazing examples – the giant beasts in Black and White, the cute critters in the mid-nineties artificial life experiment, Creatures… But can true enemy unpredictability work in games? Today’s AI coders aren’t so sure.

Participants
Owen O’Brien: Senior Producer, Mirror’s Edge, DICE
Alex Champandard: AI coder, author and blogger (http://aigamedev.com)
Jean-Christophe Capdevila: AI team leader, Alone in the Dark, Eden Games
Jeff Orkin: MIT Media Lab, Cognitive Machines Group. Previously Senior

Owen O’Brien: I think it’s nice when you play an encounter in a game in different ways and see different reactions. This doesn’t necessarily mean unpredictability though, just that they react differently to different situations. At their heart most games are about understanding and mastering patterns of input and output. You don’t want a completely unpredictable opponent since part of the enjoyment comes from learning how they respond to you and then using that knowledge to overcome them.

This goes as much for game geometry as it does for AI. However building in degrees of unpredictability can mix things up and keep them fresh. I don’t see any conflict with a narrative based game. If you need certain things to occur there is no reason why you can’t use a mixture of methods from letting the AI run rampant to tightly scripted sequences. I know I keep coming back to it but all this is so heavily dependent on the kind of game you are making and the experience you want the player to have.

Alex Champandard: Game AI has really hit its stride in the last 2-3 years, thanks to some particularly talented game developers who have taken the most promising classical AI techniques and applied them to game AI. Classical AI is basically an approach to artificial intelligence that involves reasoning with symbols, like planning. This technology has been around in some form for half a century, but it’s perfect for AI in games — and we’re rediscovering that very quickly now we have a little more processing power to leverage.

On the other hand, Modern AI including neural networks (NN) and genetic algorithms (GA) have only turned out useful in very specific applications, and have not proven themselves to be as much use in game development. For the bulk of the games industry, it’s unlikely that will change in the short term as only companies with huge budgets can experiment with such high-risk and potentially unfruitful technologies. It’s not to say that NN or GA are completely useless for game AI, but professionals do not consider them as viable solutions very often.

Game AI is fundamentally about smoke and mirrors — and that won’t change. Think of it as developing the AI for a professional actor that plays a convincing and challenging soldier, compared with implementing the soldier AI in the first place. That said, game developers still need better AI technology to make sure that the smoke and mirrors are placed intelligently! In particular, it’s important to create enemies that appear believable to the player in every detail, but leverage every other opportunity in a plausible way to make sure the player is having
fun with the game. Getting that right often requires twice the amount
of work, but it’s certainly worth it…

Louis Gascoigne: Unpredictability is extremely important. In Dead Space there was a constant battle between designing (scripting) scary versus having Necromorph behaviors that end up just being scary. Ultimately both were used.

The Necromorph’s charge behavior is random with the exception of a bit of logic, which determines whether it would be fair or not. he Necromorphs use dead reckoning to extrapolate the player’s movement to time their leaping attack and it takes off quite a bit of damage.

Jeff Orkin: Game development is a story-telling medium, and every story needs a script. Scripting won’t go away, but its role and granularity will change. As AI characters become more capable of improvisation, they will be able to take direction at a higher level, and become collaborators in goal of providing an exhilarating experience for the player.

Jean-Christophe Capdevila: The most important question is whether this is what we want? How do we design a game where the NPC thinks and perceives the same thing as the player? If the NPC learns from his errors and even those of the player, and does not repeat the same behavior, trying always to find a way to counter-attack or thwart the player, even if it means frustrating him in the end, is this what we want?

In the context of an FPS during combat, this could be interesting, but how do you build a game around a story without giving sufficient controls to the designers? Imagine a NPC refusing to execute a simple order to go straight ahead, because he “knows” (because he has already lived the situation) that the player is waiting for him and that it risks dying, when this phase is essential to the story or to the level design of the situation. This aspect would be gratifying to the programmers, but I can already see the designers asking themselves how they will be able to handle NPCs that will in the end become totally uncontrollable.

Keith Stuart