A Retrospective Analysis of Game AI in 2008

Alex J. Champandard on December 31, 2008

Last year, I wrote about my Top 5 Trends and Predictions for 2008. Not only was it a fun article to write, but it also very popular! Now seems like a good time to go back over those five predictions and see how accurate they actually were. I’ll also add a few more trends of 2008 that I didn’t predict but turned out to be very important nonetheless…

Cooperative AI

Army of Two

Screenshot 1: Cooperation in Army of Two.

Last year, I wrote this:

Prediction: Increasingly, developers with the bigger budgets and whole sub-teams of AI designers, animators and programmers will include cooperative AI characters to help emphasize the player’s emotional involvement with the story. In particular, watch out for the dog in Fable 2, which is an amazing idea and a lower-risk way to provide an AI sidekick.

This prediction was a little obvious, but rather accurate in the end! Fable 2 was just one of the examples, since cooperative AI featured in many games including Left 4 Dead, Army of Two, and to a lesser extent Grand Theft Auto 4, Gears of War 2 and Resistance 2 — among others. Early next year, games like Killzone 2 will also feature cooperative AI characters heavily. This is definitely an increase compared to previous years.

What I didn’t mention in my prediction is the difficulty of implementing cooperative AI. The AI characters in many of these games received mixed reviews from journalists and players alike. It’s certainly not easy to build an AI that performs well while constantly under scrutiny. I expect developers will continue to rely on cooperative NPC characters, and slowly get better at the process.

Sandbox Games

Grand Theft Auto 4

Screenshot 2: Grand Theft Auto 4’s trademark sandbox.

Here’s what I wrote last year on the subject:

Prediction: GTA 4 will take the genre to new levels with its next-gen AI, but sandboxes will generally become more popular as they provide diverse gameplay and more replayability. These environments will allow the AI to show it’s true colors… Expect to see both amazingly intelligent or terribly stupid enemies as developers try to move away from short-term scripting!

Grand Theft Auto 4 did not disappoint in this area, but many other games were built around this sandbox concept that features AI heavily. In particular, Saints Row 2 and Mercenaries 2 applied these ideas to similar settings. Other games like Resistance 2 (multiplayer) and Left 4 Dead implemented sandbox-like gameplay in a more focused fashion.

The sandbox setting certainly revealed its fair share of AI bugs — as predicted. The AI also helped improve replayability in most cases, within the limit of the game’s design and original appeal. I doubt the trend of sandbox games will increase much more (particularly with publishers cutting development budgets), but expect to see more AI characters designed to be autonomous parts of a simulation…

Emergent Behaviors within Stories

Screenshot 3: Far Cry 2 slightly less linear gameplay.

Last year, I predicted the following:

Prediction: Even in story-driven games, developers will increasingly use pockets of emergent behaviors to increase variety and replayability. This is a great way to bring a sandbox-like environment into a more traditional AAA form, getting the benefits of both.

I wrote that with BioShock in mind, but a few other games this year attempted to take these ideas further. Far Cry 2 got a lot of press for attempting to bring back the sandbox into the first-person shooter genre, since Crysis turned out to be a little more linear than its predecessor. In particular, Ubisoft’s title had elements of simulation (such as fire and wildlife) that brought an extra level of non-linearity to the game itself.

S.T.A.L.K.E.R. Clear Sky received a little less focus from the press, but took the concepts from Shadow of Chernobyl to the next level. In particular, the game by GSC Game World now models the behavior at a faction level, so each group of Stalkers is constantly assessing which part of the land to take over and figuring out how to achieve these objectives — like a small game of Risk. (AiGameDev.com members can see the full interview video; more information here for everyone else.)

There’s still a long way to go along this trend, as it’s by far the least mature area of game AI — but the early results are encouraging.

Hierarchical Planners

I was a little over-zealous in predicting the following for 2008:

Prediction: In 2008, expect to see even more planning in games as well as middleware vendors for game AI. This should not only help improve the development efficiency so more time can be spent on little details, but it should also make the behaviors more intelligent. Many of the features discussed in the previous points will become possible thanks to this kind of technology.

Behavior Tree

If you consider the term “hierarchical planner” to imply a Hierarchical Task Network implementation, then this prediction didn’t happen as I described it. No middleware vendors have yet shipped planning solutions based on HTN, and no games have been announced to be using HTN either… However, I’m going to stick to this prediction for the mid-term future, and hope that things in the pipeline move a little faster!

What did happen during 2008 was much closer to my informal prediction in this comment:

“A great boost in popularity of behavior trees and their editors.”

Here at AiGameDev.com, I’ve been evangelizing behavior trees for over a year and a half now, and over the past few months I’ve been consulting with a handful of AAA studios about behavior trees and their implementations. But in a sense, I completely underestimated the influence of the blog itself as well as the strength of behavior trees in general. It’s now safe to say that BTs have become the de facto standard in the games industry for anyone writing a new system from scratch.

Further Information:

Scripting Languages


Screenshot 4: Scripted sequences in Crysis.

This was the last item from last year’s list:

Prediction: Rather obviously, 2008 will see an even wider adoption of Lua from across the spectrum of AAA games to independent games. Developers are finally getting a grip on building good debuggers for the language, handling errors robustly, and getting comfortable with its semi-coroutines. However, as companies use the language for more and more aspects of the game logic and AI, some cracks will show up increasingly.

While not very risky, these predictions turned out to be rather accurate also. For scripting, Lua is still dominant — although competing implementations are gaining traction. For instance, according to Lead Programmer Dmitriy Iassenev, S.T.A.L.K.E.R. uses LuaJIT — which works on the PC platform only.

From the AI perspective, developers which previously offered Lua as a solution are increasingly switching to other techniques such as behavior trees, which can be implemented at a much lower level in C++ and optimized significantly compared to byte-code interpreters. For example, Crytek’s Ricardo Pillosu recently did a talk about behavior coordination between agents using behavior trees (talk overview in Spanish). For reference, the CryEngine’s AI was formerly a Lua poster child.

AI as Procedural Content Generator


Screenshot 5: Spore’s procedurally-generated landscape content.

Now for an increasing trend that I only covered implicitly last year: procedural content. Games like Spore in particular have emphasized the benefits of this approach, both as a way to foster unique gameplay as well as a tool for reducing content creation times. For a few years now in academia, AI has fallen out of favor in terms of funding, but with renewed interest in procedural methods AI is finding new ways to wiggle into research projects — since the technology is particularly suited to help with content generation in a variety of ways.

That said, over in industry, game AI on its own has benefited from an increased focus over the last few years, so this procedural trend is yet another way for AI programmers to have an impact on modern game development! This has been obvious with various papers published this year, for instance this paper about Spore’s animation system (although it requires a lot of animation data rather than procedural code than people expected).

Increased Awareness and Coordination

Paris Game AI Workshop 2008

Screenshot 6: Lunch at our Paris Game AI Workshop in 2008.

Another trend that established itself this year is that the game AI community as a whole is becoming more organized! This is partly due to an increasing focus being placed on AI in the games industry, but also because people around the world have decided to take action this year. Among others:

  • The AI Games Network was founded in the U.K. this year with EPSRC funding, with the goal of helping academia connect to industry. (If you’re attending the event in Bradford in January, let me know!)

  • Our dedicated Paris Game AI Workshop (photo above) was held in June this year around another game conference, and was sold out! Based on the success we’re holding another 2-day conference on June 10th & 11th in 2009 too.

  • There’s a new IEEE journal called Transactions on Computational Intelligence and AI in Games, which is sounding promising from the papers I reviewed so far. The first issue is scheduled for March.

  • The AI Game Programmer’s Guild was also launched in stealth-mode this year, including myself and select others from the forums here. The result is an extra two days of AI sessions at GDC this year!

  • Last but not least, AiGameDev.com has been growing very healthily; in October we launched our continuous training program for game AI professionals, which has received great feedback so far. (The program will be re-opened for new members in January; stay tuned.)

From this perspective, 2008 has paved the road for a very prosperous few years ahead!

Discussion 0 Comments

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!