Interview with Ramon Axelrod, CTO of AIseek

April 18, 2007

This week we continue our series of interviews with the Online Game Development Conference's high-caliber speakers and move into a more technical direction with Ramon Axelrod, CTO of AIseek. Ramon will be giving a great lecture at OGDC, Intelligence for Changing Worlds: Real-time Generation of AI Data in Dynamic MMO Environments. Ramon, who holds a B.Sc. in Mathematics and Physics, and has a background in advanced military simulations, discusses why he thinks that AI is holding back game design.

Edward Van Duering

Edward Van Duering: Thank you for letting us steal a bit of your time, Ramon, to have a few questions answered. AIseek may be a new name to a lot of folks, so perhaps you can tell us what your company does and what you hope to accomplish in the game industry.

Ramon Axelrod: Our industry is all about innovation — creating the next great game that will thrill players, outshine the competition, and create a winning franchise. This content innovation is very much driven by technology. And that’s where AI comes in. I believe that a leap forward in AI capabilities will enable a new wave of game innovation. However, such a breakthrough requires a lot more computational power dedicated to AI — just as graphics acceleration hardware made possible the 3D worlds of today. AIseek is making this happen by introducing an "accelerated AI" solution — an AI library backed by the first dedicated processor for game AI. For MMOs, we’re making our AI chip available via an add-on card for game servers.

In your estimation, are current AI techniques ahead of or behind the curve of game design in general?

Over the years AI programmers have developed some very advanced AI techniques, but major leaps forward have been rare. The fact that other areas of gaming have moved faster makes it look as though AI is lagging behind. Take the example of pathfinding. For years, we’ve thought that this was a solved problem. But today, with game worlds becoming much more dynamic, even something as basic as pathfinding has to be revisited – or maybe overhauled. It doesn’t have to be this way. If you could build the entire AI graph from the game’s geometry whenever you wished, and compute paths fast enough, pathfinding would be easy.

So, it’s not just that AI is lagging behind — you might argue that it’s actually holding back game design as a whole. This is what I call the "NPC trade-off". Due to limited computing resources for AI, game designers and AI programmers are forced to choose between three competing parameters: the number of NPCs (non-player characters), the intelligence of those NPCs, and how dynamic the game world is (how much data needs to be re-computed on-the-fly). So you compromise. You design your game so as to require fewer NPCs, or only very simple NPC behaviors, or maintain a largely static world (or some combination of these!) I think the right amount of processing power for AI can help close the gap between what's possible today in other areas of the game and what's possible in AI.

Which MMO or online-enabled games, regardless of platform, are the best examples of what good AI should be like in a game?

I think that MMOs are only beginning to realize the potential of better AI to create unique gameplay experiences. They are looking over their shoulder at what's being done on the client-side, and thinking "How can I get some of those intelligent behaviors into my game? How can I make my game world physically dynamic? And if I do that, how will my NPCs survive?” That's exciting for players, but it creates a huge challenge for AI.

Stepping into the 'what if" realm, present an example of ideal behavior be of an NPC in a given game? Could what you described actually be done now or is it still off in the future?

Let's imagine a fantasy MMORPG, only much more dynamic than what we are used to today. In this MMO, players and NPCs can produce real changes in the world, and NPCs are expected to be much more active in the game. They interact with the environment by moving objects, thus creating new passages, blocking existing ones, and finding new places for taking cover. NPCs also search through many more options to avoid the obvious pitfalls. I believe that such scenarios can be realized today, given some clever programming and enough AI processing power. Looking further into the future, I see how certain technologies for action planning and body articulation (now in their infancy) will mature and enable a wider range of interactions.

To program a truly effective AI algorithm in a complex game, to a development staff how time-intensive and resource-intensive is it? Is that additional programming worth the extra effort?

Many of the best techniques are well documented in the game AI literature. Nevertheless, implementing such techniques in a complex game can take a couple of talented AI programmers many months. There’s no question that it’s worth the effort — stupid NPC behaviors can be very disruptive to the game experience and are a big turn-off for players.

Are software or hardware-based solutions the best path to enhancing AI in a game?

There's no doubt that traditional, software-based AI (running on the CPU) is getting better and better. In fact, while AIseek accelerates the bulk of the AI on our Intia AI processor, we still take advantage of the CPU for certain computations. So I think that if you are looking for a modest increase in AI scope and quality, sticking with a software-only approach makes sense. On the other hand, if you're looking for a very significant leap forward — something that really drives new gameplay — then the hardware-accelerated option is worth a serious look.

If you would, give us a taste of the new information and ideas you plan to present for the attendees at the Online Game Development Conference.

MMO designers are beginning to introduce dynamic changes into their game worlds. So keeping the AI up-to-date is now a real challenge. To address this, I’ll be talking about a new technique for updating AI navigation graphs in real-time. The technique starts with the game’s raw geometry – what is sometimes called the “polygon soup” — and processes this on the CPU. The result is that the AI graph is generated and kept up-to-date automatically. I’ll also show how this approach can be accelerated using the GPU to run part of the algorithm.

What can we expect from AISeek during the remainder of 2007?

Over the next few months, we'll be releasing our Dev Kit for accelerated AI to the first set of MMO early adopters. The Dev Kit includes both an SDK for integration with the game engine and our AI acceleration cards for MMO servers. This means MMO developers can experience the performance gains first-hand and begin to implement the new gameplay possibilities. So the remainder of 2007 should be a very exciting year for accelerated AI, and for MMOs in particular.

Free Newsletter

Join the OGDC update list and receive our free weekly newsletter with all the latest conference and online games industry news.

Premier Sponsors