Open Editorial

This Year in Game AI: Analysis, Trends from 2010 and Predictions for 2011

Alex J. Champandard on January 4, 2011

Every year, I write an analysis of the past year's trends and break down predictions from 12 months ago to see how accurate they were. Without fail people ask me whether much has happened during the last 12 months in game AI. But after spending a few weeks going through news and events, I always come to the conclusion there's way more than meets the eye! Last year was quite an incredible year for artificial intelligence in games, and this year is living up to those high standards.

As for the predictions themselves, there's always been a certain self-fulfilling aspect to these posts, because we obviously remain very active working on all of these topics at throughout the year. However, until recently I under-estimated the influence of the blog and its reach, which turns my thoughtful rambling analyses into occasional points of reference for the community. Rest assured though, I won't let the pressure get to me and I promise to be just as controversial this time around!

Editor's Note: I included many links and references to the various sources for the trends, predictions and analyses. Some of those are for PREMIUM-level access only, but fear not; there's no better time to join! Not only do you get a whole bonus month for joining before January 15th (during our winter break) and our introductory price has never been more attractive... Now's your chance to follow up on those new year resolutions. *wink*

7 Prediction in Retrospect

At the start of 2010, I wrote some predictions for the year in the Game AI Forums right here on I'll summarize them here and post an analysis to see how close things turned out.

Prediction #1
Production Maturity

“Based on topics and recent discussions like methodical testing of AI (i.e. Matthew Jack's masterclass, PREMIUM) and the use of determinism to help isolate bugs or create predictable save games, I think we'll see steady progress in the area of software production for AI this year. It will mean less time is spent on isolating hard bugs that are hard to reproduce, and a much tighter integration of programmers and their Q&A team.”

Technology in the games industry has indeed matured, particularly as developers master this generation of consoles. Now their infrastructure is in place, more studios are focusing on debugging tools. For example, this includes behavior tree editors and general AI tools for capturing events and generating traces from them. See for instance the tool they're using at IO Interactive from Mika Vehkala's masterclass (PREMIUM).

It's not been a radical jump, but the "steady progress" I predicted is certainly noticeable. That said, there's not much to report the topic of determinism and its applications to the Quality Assurance process. Apart from the usual RTS games that rely on determinism for networking, pretty much only HALO: REACH has used this approach — though rumours have surfaced that BRÜTAL LEGEND did something similar too the year before.

Prediction #2
No More FLOPS for You!

“Despite the various multi-threading techniques and toolkits available (e.g. TBB, OpenCL) designers will realize you can't use any of the "optional" computation power (above the basic requirements) for AI without completely changing the gameplay. For instance, if you have 1 TFLOPS to spare, putting that into the AI will make it behave differently (e.g. smarter, trickier, etc.) and will require you balancing that game completely separately as a new SKU. Consoles of course, don't have that problem as you have a fixed amount of computation.

This will lead to developers trying to build fixed AI/gameplay, but try to portray it using more computation when it's available. For instance, better crowd animations, or using GPU to simulate large groups, but according to the simple patterns that remain the same on everyone's hardware. Borut Pfeifer was pointing in this direction during his interview on I think this trend will continue.”

Generally speaking, GPGPUs (general purpose GPUs) had a slightly disapointing year in terms of game AI or game simulation. Granted, there have been some interesting research and prototypes in computer graphics, but not many notable commercial applications for gameplay. We have yet to see a killer application that does something we can't already do similarly with equally optimized CPU code. In fact, the big GPU manufacturers seem most concerned with high-performance computing (HPC) market rather than gamers, including Intel refocusing its Larrabee efforts and NVIDIA's revenue stream apparently being dominated by HPC applications for its new Fermi chip.

In terms of games using more computation to portray scenes better without affecting the gameplay too heavily, HEAVY RAIN came the closest with its crowd simulation. At the Paris Game/AI Conference 2010 in June, the team discussed their solution — which was based on the Continuum Crowds demo, and the frobblins demo. WIPEOUT HD also featured ambient animation in the background to make each racing track much richer visually, as did GEARS OF WAR 2 — though both of those were from two years ago.

Prediction #3
The Return of the Bot

“Playing online isn't for everyone; it's a stressful and sometimes unnecessarily emotional affair. Offline you can play against bots without pressure and invite a few friends maybe, to get most of the benefits of the game's design and sometimes more. (Bots always behave like gentlemen.) The golden age of bots was with highly moddable PC games like Quake I/II/III. With games like GEARS OF WAR 2 and KILLZONE 2, bots are making their way into mainstream console games.

I think this trend will continue. As game developers or publishers try to make good online components to attract and keep players attention, there simply won't be enough people to play every game. (Most will be playing L4D2, HALO REACH, or possibly MW2). Bots will provide a good way to maintain a good experience for players even if the online community isn't as big as these AAA games. This was the case with DEMIGOD.”

There are mixed feelings about bots generally, and they surfaced this year. The community has a love-hate relationship with bots — as we discussed in a technical interview with Jeremy Swigart, author of the omni-bot. Bots are often blamed for destroying game communities (e.g. WOLFENSTEIN: ENEMY TERRITORY), as you tend to find more servers with bots in them as a game gets older. This creates a lot of hate towards bots, though the communities would die down eventually anyway (and probably quicker) without bots.

However, as I pointed out in my talk at GDC Europe, the benefits of bots outweight the cost — as they fix many experience shattering issues with online play. Valve in particular is taking its bot technology from LEFT 4 DEAD further, and they were integrated into TF2 towards the start of 2010 as an experiment. You can now play with bots as an offline training mode from the menu in TF2, and if you haven't played online before the experience easily compares to random games against hardcore players in unbalanced teams!

Prediction #4
Motion Control for Dummies

“A big trend I anticipate for this year, assuming Microsoft and Sony ship their new peripherals on schedule, is motion controllers. What we'll see is that game developers are always very pragmatic about using the new technology, very much like on Wii. So instead of relying on tricky machine learning, they'll just put together simple hand-coded recognizers to find motions in the patterns.

In a way, the design of the new motion controllers (e.g. Sony's 3D tracking, or the Wii Plus) come from weaknesses in the ability to induce motion from imprecise sensors, so adding more sensors and keeping the code dumb seems to be the way forward. I expect to have this confirmed in 2010.”

Motion control was indeed a huge part of 2010, and it was Microsoft's Kinect that captured the imagination of the community — especially after the USB protocol was 'reverse engineered' and open-source drivers were made available. Kinect is based on an innovative camera that captures colored images as well as depth information. There's certainly some advanced software technology under the hood, hidden within the device itself, but that's not what stole the show...

Over the past months, the multitude of amateur experiments with the device became increasingly creative, ranging from puppeteering to gesture recognition and World of Warcraft controllers. As I predicted, many of these experiments were indeed done with simple recognizers designed specifically for those applications. However, in most cases, computer vision libraries under the hood provided information about the position of skeletons in space, making these home-brew hacks much more robust and sophisticated. By not encrypting Kinect's USB connection and making the device relatively affordable compared to alternative RGB+D cameras, Microsoft has guaranteed a strong community of developers and this trend will most likely continue to grow from strength to strength in 2011.

Prediction #5
Not More Behavior Trees?

“I think we've reached critical mass with behavior trees now. BTs are awesome of course, but I think some developers are just using the word now because it's popular and cool. A "BT" isn't always the best way to describe your architecture, and this has been the case in a few games for 2009 already...

For 2010, I expect a certain lassitude with the topic. Because of this, I expect developers will start to refine their understanding better, so they can differentiate HFSM and BTs better, and talk about what kinds of customizations they're using to differentiate their architecture.”

It's been a very good year for behavior trees, and there was less lassitude than I feared. Not only have more AAA studios shifted to BTs than ever before, but there are now middleware solutions starting to become available and even a few open source implementations (in early stage) in C++ and C# too. More importantly, however, the community is increasingly standardizing its terminology. In my GDC Europe 2007 talk entitled Behavior Trees for Next-Gen Game AI, I introduced some terminology to describe nodes in a behavior tree (priority selectors, random selectors, sequences, decorators, filters, etc.) an those terms have become commonplace now. I also introduced a graphical notation for BTs borrowed from HTN planner literature, and that has also picked up momentum as a great way to represent and explain these behavior trees.

Prediction #6
Rediscovering Old-School Game AI

“With more and more games coming out for iPhone, DS, PSP and Flash, there'll be more emphasis on old-school AI techniques that we used years ago, and almost a rebirth of these techniques. I'm thinking about steering and non-hierarchical state machines in particular.

I'm hoping for 2010 more of these indie / portable platform developers go into more modern AI, and find that doesn't in fact require that much processing power either — but can affect your gameplay tremendously.”

Obviously there have been many social and mobile games released this year, and the AI has certainly got the job done using whatever technique the developers felt more comfortable with. These studios rarely disclose what kind of AI techniques they use though, but finite-state machines and scripts are a good place to start. There has also been more discussion of AI techniques for indie games, in particular during a group lecture at the AI Summit 2010 earlier this year; Phil Carlisle pointed out that putting smart objects into the world, along with level triggers and their corresponding scripts, is a very effective way to work.

In terms of old-school AI in modern games, there was an in-depth analysis of PACMAN recently, including the behavior of its ghosts. The patterns of attack-retreat, as well as the emergent chasing down of the player from simple behaviors, are surprisingly reminiscent of LEFT 4 DEAD's zombie mechanics and its procedural director.

Prediction #7
Animation and Motion Capture Repositories

“Smaller game developers spend a lot of time simply re-animating or re-capturing animations that are very common, such as walk cycles, run cycles, etc. Sites like Mixamo and Animeeple can in principle step-in to help with this problem by providing a shared repository. In the case of Mixamo, There are technical and financial roadblocks though, as parametric animations are tough to reproduce client side, and the cost of the animations is high given the quality.”

This trend actually died down in terms of news and usage... If anything, it seems AAA studios are more interested in building their own internal motion capture studios, for example as Bungie did for HALO: REACH. With motion capture hardware (in its various configurations and devices) becoming so cheap, this seems to be a more reliable alternative for studios that would require cut-scene animations anyway.

I was holding out for an open/shared repository of motion capture that can be used by game developers at low cost and without royalties, but in practice the case for simple procedural animation makes much more sense for independent and low-budget studios. So while animation websites and databases have not hit their stride yet, there's only a small window of opportunity for them to do so before procedural solutions become available. Based on the reception of Ken Perlin's talk at the Paris Game/AI Conference, such procedural animation techniques will be becoming more popular as more people integrate them into their games.

10 Unexpected Trends

Here are some trends that turned out to be very important this year, but nobody really predicted...

Trend #8
Post-Release AI Polishing

When you're building a complex product like a strategy game, it's hard to build final quality AI before the game has shipped. So many things are changing in the game's engine and the design itself that you're always aiming towards a moving target. Even when things stabilize, for example moving into beta, you probably won't be making any radical changes anyway to keep things as stable as possible before shipping... This goes a long way towards explaining the patches that address the initially released AI.

In 2010, we've seen a wide variety of patches that fix and significantly improve the AI. Games like CIVILIZATION 5, SUPREME COMMANDER 2, AI WAR, EMPIRE: TOTAL WAR all have seen patches with major changes to the AI. For example, SUPREME COMMANDER 2 saw multiple patches with a significant time investment from Mike Robbins and the QA team (see the interview with Mike Robbins). The guys at Arcen behind AI WAR also put their faith into the game's longevity by releasing many free iterations, and are now up to version 4.0!

While there are some commercial benefits to polishing a game's AI post-release, it's not going to change sales statistics dramatically. The AI WAR team ran into some financial trouble by relying too heavily on continuing sales. In the case of SUPREME COMMANDER 2, there's a deliberate plan to reuse the technology on the next project and use the comfort of a stable project to polish it. If you're planning to improve the AI of your game post release, do so strategically with your next title in mind, and consider combining the patch with a sale, promotion or marketing push.

Trend #9
Investing in AI

Not every single studio takes AI seriously, but those that do are increasingly becoming more strategic about it and investing in their codebase. This ties in to the previous trend about polishing the AI post-release, but taking it one step further:

  • Valve started the year by introducing bots into TEAM FORTRESS 2 as part of their initiative to develop their AI from LEFT 4 DEAD further. Valve often uses TF2 and its community as a proving ground for various experiments, and obviously was a perfect fit for prototyping bot AI that can achieve in-game objectives and deal with multiple abilities.

  • The Creative Assembly has been building up the AI for their Total War series, as I mentioned both their EMPIRE and NAPOLEON games saw many iterations and improvements. Beyond that, the company has made the commitment to get the AI right for SHOGUN 2 at release. Having built up and invested in the technology over multiple games makes this possible.

  • Similarly, Gas Powered Games invested in its AI from SUPREME COMMANDER 2 over the course of multiple patches and many months of work. The company is now reusing its technology as part of its next game, KINGS AND CASTLES.

Of course, the major engine providers are also building up their AI technology — and so far they seem to be doing this in-house rather than through partnerships or acquisitions.

Trend #10
GOAP Is Dead. Long Live...

Earlier this year, I was on a panel called "The Right Tool for the Job" where I represented planners. I was very torn about that at the time, since planners cover a very wide spectrum from STRIPS-style goal-oriented action planning all the way to hierarchical task network planning — and those are very different in practice. Now, here's where I'm supposed to say that AI techniques are just tools, and different problems require different tools. But the fact is that building non-player character AI for games is already a very specific problem, and STRIPS-style goal oriented action planning has turned out to be a dead end. After reaching a peak of adoption a few years after Jeff Orkin's seminal talk at GDC 2006, more studios than ever have switched away from the STRIPS-style GOAP that the original F.E.A.R. used.

In particular, Avalanche switched away from GOAP after JUST CAUSE 2, DEMIGOD's hybrid planner wasn't reused in SUPREME COMMANDER 2 at Gas Powered Games (it's now a neural network, of all things) and METRO 2033's original STRIPS implementation was replaced with a prioritized list of complex hand-crafted behaviors. The many issues with STRIPS are well documented, in particular that it requires you to add unnecessary variables to the world representation just to make it understand simple concepts, and that the only other way you really have to control the behavior is using weights & heuristics which also affect performance tremendously. I call this "baby-sitting" the planner, and it's a tricky balancing job to say the least.

Academia has long discarded such planners in favor of hierarchical ones, particularly when performance matters. Obviously for games it does too, but you can also control hierarchical planners much better thanks to their top-down task decomposition mechanisms. If you're thinking to yourself, Guerrilla Games saw this coming and were a step ahead with KILLZONE 2, then you're absolutely right (see this panel discussion at the Paris Game/AI Conference 2009). However, while GOAP has turned out to be a dead end, it's not necessarily the case that HTN planners will shine from now on. Unfortunately for planners, there's too much similarity between behavior trees and HTN planners, so it's much harder to justify the extra implementation time during development and computational effort at runtime.

Trend #11
Terminology Standardization

I already mentioned that the community is coming to a consensus on terminology for behavior trees, but that's not the only domain where we're standardizing our understanding. Utility systems have also become a common term in forums, at conferences and occasional discussions. However, most importantly, we've converged towards the label of computational behavior to describe the overall field of artificial intelligence for games (as a result of discussions with Phil Carlisle and Mikko Mononen). We feel that "game AI" is a little misleading to beginners, and doesn't help researchers in the field who are coming from a traditional AI background.

Calling it computational behavior should help a lot in all those domains! It's not about intelligence for its own sake, it's all about the results in game. In that sense, what we used to think of "game AI" is completely at the opposite of the spectrum of "artificial general intelligence" which instead models intelligence and thought patterns in the brain for their own sake — hoping (at best) they emerge into cool behaviors. Game developers can leverage that kind of emergence too, but it will be much more deliberate than the way AGI research does it. Overall, the term "computational behavior" reflects the additional insights that game developers are now bringing to the field, taking ideas of robotics further and very selectively borrowing AI or machine learning techniques.

On that subject...

Trend #12
Polarization of Academia

Every year without fail on, we discuss the huge gap between academic research and the games industry... but rarely are those conversations very productive. This year, I spent more time than ever talking with researchers, giving presentations at academic conferences, networking and getting their feedback on the topic. Today, as far as I'm concerned, I'm going to proclaim: there is no longer a gap between academia and the games industry! What's happened is that the academics who wanted to bridge that gap have done so, and this year it's particularly noticeable. There was a handful of very relevant talks and papers at CIG 2010, and AIIDE 2010 had a much higher average of applicable research. Many people from industry have claimed that there's sufficient information out there for anyone to help bridge the gap thanks to books, conferences, networking opportunities, as well as itself — and that's proven to be the case.

What's become apparent is that not all academics seem to want to bridge the gap... What we've ended up with is 10%-20% of the academic research community in AI & games working on projects that are relevant to the games industry. Then, based on my estimates, there's a large remaining majority that's instead interested in applying AI to games as a test bed; this group more interested in the AI for its own sake rather than solving the problems that games industry is facing. In fact, this group often solves problems from other industries without discrimination and has no particular interest nor passion like game developers tend to have.

Obviously, there's still lots to do to make sure the 20% does even more relevant work and gets even better feedback. That's a big priority for us, and for instance, our call for proposals for our Paris Game/AI Conference 2011 is designed to address that; you don't have to submit a full paper and you'll get quick feedback on your research project. As for the remaining 80% of academics, they tend to be a bit more vocal! What they want is pre-prepared datasets and easy-to-use maintained source code that they can use to evaluate their own theories. Unfortunately for most game companies, there's no measurable short-term gain in preparing these datasets or easy-to-use SDKs, but there's certainly an opportunity for an entreprising developer to release a tough problem and end up with a crowd-sourced solution that cost nothing.

Trend #13
Flow-Field Pathfinding

A few years after the publication of convincing research about flow-fields for crowd simulation, specifically Continuum Crowds and the March of the Froblins, this technology found its way into multiple games from 2010.

  • SUPREME COMMANDER 2 is using flow-fields on all of its units for pathfinding, which allows multiple large platoons heading in opposite directions to cross over in the same place and form lanes. According to Gas Powered Games' video blog, a hierarchical version of this flow-field algorithm is under development for KINGS & CASTLES.

  • HEAVY RAIN uses a SPU-optimized version of the Froblins algorithm to simulate its crowds in some special scenes from the game. This algorithm is used to replace its traditional A* implementation in those special cases, to make sure the crowds looked convincing enough.

  • STARCRAFT 2 seems to use technology with a very similar effect for large groups, though it's not yet clear if this technology is applied consistently to all units. This is particularly obvious for large armies of Zerglings (400) with additional speed, as you can see the last units litterally "pour" into the target location and push the others out.

The down side of this technology is that the behavior of individual units is lower quality when you can isolate them. In practice, STARCRAFT 2 manages this expectation best because you don't expect much from the abilities of individual Zerglings! What's also fascinating is that it all happened within a relatively narrow window of time, plus you can measure gap betweeen publication of the research and its appearance within games.

Trend #14
Middleware Explosions

Two years ago, I predicted a consolidation of middleware companies in game AI. That's partially happened since then, but it's been more of an explosion into different directions than consolidation. Basically, I under-estimated the strength of related markets like serious games or simulations which seems to have helped supplant the income of many middleware vendors during tough times in the games industry. For instance, Spir.Ops has refocused its efforts on private R&D for games and simulation projects; AI.implant was bought and integrated into Presagis Inc. and seems to have changed its direction accordingly; TruSoft's core team has moved on and its CTO Iskander Umarov is now Senior AI Engineer at Irrational Games.

The general middleware business this year was shocked by the announcement that Emergent Game Technologies was effectively closing down and auctioning off its assets, including their flagship game engine GameBryo. Based on a blog post by Vincent Schieb, formerly Software Architect at Emergent, the cause seems to be a mismatch between what the investors expected and what the business could deliver (notice it was never profitable given those high expectations). This signals to me that AI middleware companies taking external funds will come under similar pressure from their investors. In particular, this is the case for Xaitment, which a few months ago announced an additional $1.6 million in a second round of funding, on top of the $1.3 million from the previous round.

In practice, I think this will cause a visible rift in the AI middleware industry as the market for navigation middleware continues to mature. The products created by established companies with solid foundations and large internal budgets will continue to grow and improve in a determined fashion. As for independent AI middleware companies, they will be under more pressure to promote their solutions by distinguishing themselves technologically and/or marketing more aggressively. Either way, hopefully there won't be too much astroturfing in forums!

Trend #15
Character-based Experiences

Phil Carlisle actually predicted the trend of featuring gameplay based on prominent characters in games, but he does that every year and not everyone listens to him. *grin* This year, he was right though! More games than before, including two very high-profile games, have relied on character-based storytelling for gameplay — which obviously includes a significant amount of AI. Those games were ALAN WAKE and HEAVY RAIN, not to mention the handful of other games with more traditional gameplay that had an increasing focus on characters. In particular, MASS EFFECT 2, ASSASSIN'S CREED: BROTHERHOOD or RED DEAD REDEMPTION.

It's clear that believable non-player characters will continue to play a prominent role in games, even if the gameplay does not revolve around them directly. Game developers are becoming more skilled at portraying such characters with current hardware, so this will no doubt be a trend for 2011 also. (Watch out for the facial motion capture in Rockstar's L.A. NOIRE particularly.) However, it's unclear if there's enough demand from gamers for story-heavy games like HEAVY RAIN or ALAN WAKE. Only one of the two was a commercial success (the PS3 title), whereas the other was one of the most pirated XBox360 games of 2010 according to recent statistics.

Trend #16
Leveraging Crowds

Now developers are comfortable with the technology for rendering dense crowds efficiently on modern consoles, we're seeing more games use crowds as a tool to enhance the experience. This year, however, it's not been for visual purposes and "ambient animation" like I wrote about last year — but for gameplay instead. In particular:

  • HEAVY RAIN leveraging the crowd in the train station or mall for example, to implement the gameplay for key scenes in the story.

  • CALL OF DUTY two years ago used a crowd in an airport controversially, but the sequel is relying on large numbers of people more often.

  • ASSASSIN'S CREED introduced elements of social stealth from the beginning, and is continuously developing gameplay around those crowds.

  • TRON: EVOLUTION also includes scenes with many NPCs in it — but I haven't played it yet!

Expect big studios to continue to leverage crowds in the future, particularly as middleware steps in to provide the technology.

Trend #17
Practical Competition

In 2010, I was impressed not only by the amount of competitions relating to AI & games, but the way they stimulated practical and creative solutions from the contestants. In particular:

  • The STARCRAFT 1 Competition at AIIDE 2010 brought together dozens of researchers, groups and institutes worldwide in four competitions starting from unit micro-management all the way to full RTS games with "macro" gameplay. This event was so popular that it stole some attention from previously successful contests like the BOTPRIZE.

  • The PLANETWARS Competition was supported by Google, and captured the attention of thousands of students, enthusiasts and hobbyist from around the internet. The winner of the contest came up with a very interesting solution to calculate the Nash equilibria written in Lisp!

  • CIG 2010 also witness a wide variety of little contests, ranging from procedural generation of Mario levels and Starcraft maps, TORCS racing car controllers, all the way to PACMAN gameplay using an API or screen capture, not forgetting the famous BOTPRIZE. All of these were fun to watch and were solved with very practical and industry-friendly techniques most often!

We think this kind of contest is a great way to build interest in the field... As a matter of fact, we sponsored the PlanetWars contest with prizes. But unfortunately for us, Jeff Cameron (the organizer) never got round to updating the website and we ended up providing the prizes without ever being mentioned once... Worst sponsorship deal ever? Regardless, we hope to continue to sponsor such contests in the future with prizes, so if you want more information just email us at <support at> and we'll see what we can do!

Expectations for 2011

Expectation A.
In-Sourcing... for AI?

This year there was a huge trend in the AAA games industry which I'm surprised has received so little attention so far. Let's call it in-sourcing. The idea is that many internal studios of large publishers work together to finish ambitious titles. This ranges from scheduling entire level-design or art teams to work on sections of levels, to programmers helping with significant features in the code. You can easily check for in-sourced games by checking the credits; some of those high-profile games include MEDAL OF HONOR, RED DEAD REDEMPTION, NEED FOR SPEED or even KILLZONE 2 a few years ago.

My prediction is that this will become more important for AI as well. Based on the trends I mentioned earlier of investing in AI, it makes even more sense to reuse the experience and even code snippets accross projects. A good way to make this happen is to assign experienced AI programmers from different studios to help out with the AI of new projects, thanks to in-sourcing. It'll be interesting to see which techniques end up spreading thanks to this and which remain stationary in single studios...

Expectation B.
Online AI for Browser & Multiplayer Games

As online browser-based games mature and become more interesting, there'll be an increasing demand for middleware solutions to provide the backend AI. This type of middleware will require the AI to be built to serve many different online games in a more efficient fashion. Beyond just low-level technology, this requires an architecture to split the AI from the game as well as an efficient interface between the two.

There's already an example of this, with BigPoint having recently licensed EKI One for its POISONVILLE online action game. This trend will most likely continue in 2011, not to mention some high-profile MMO scheduled for released this year as well.

Expectation C.
Interactive Films

Slightly over a month ago, CALL OF DUTY: BLACK OPS was released. The controversy this year revolved around the fact that you can effectively complete part of the game without really participating! Some sections on rails you don't need to shoot at all and the worse that can happen are on-screen damage effects. In other parts of the game, if you use cover well enough you can finish the section and move on to the next! This was made entirely possible using a combination of scripting and character AI, to make sure your squad mates can deal with the waves of enemies without you interfering.

Obviously, with the recent success of replay-heavy games such as VVVVVV or SUPER MEAT BOY, this caused a stir among hardcore gamers who seemed to expect their story-driven experiences to play more like an old-school platformer... But when you keep on failing and have to replay a section of the game over and over, including the cutscenes, it ruins any sense of immersion in the story and experience. So in a sense, the "interactive film" model of COD: BLOPS is a step forward.

Where does AI fit in? The challenge is to make an interactive film that reacts more to what the player does. If you play an on-rails section badly and keep missing your targets, then the game should respond to that by for examples getting your squad mates to point out your bad performance or start mocking you. If you play a section of the game only taking cover, the next in-game scene would mention your cowardly behavior! Think of it as bringing elements of HEAVY RAIN and how it deals with player input (i.e. no replay, actions have consequences) into other story-based games.

Expectation D.
Beyond Façade

On multiple occasions this year, I had the pleasure of talking with Michael Mateas generally and about Façade specifically. At CIG, Michael's students joked how long it takes before any AI conversation drifts towards the design and implementation of Façade. I estimate that to be around 17 minutes :) But think of it as a sign of how much impact the game has had on the community and why I put it in the list of most influencial AI games. If you haven't yet studied the technology behind the game, you should definitely do so!

The game did a lot of new things in certain important and promising areas (AI and drama management), but at the same time fell short in other aspects (gameplay and story) which made it closer to a technical prototype that most people only played for a few minutes on average. I personally felt rather frustrated to listen and interact with the characters, Trip & Grace, for any duration of time — although judging them on that level is a sign of how well the characters were implemented.

Anyway, this is the year for the community to get over Façade! Now it's time for us to build on those ideas and take them further. Michael Mateas and his research group are working on another game, The Prom, which explores gameplay based on social dynamics — and based on early demos it seems to have broad appeal. Another good candidate is Chris Hecker's SpyParty, where players must pretent to be NPCs to avoid being detected as a spy during a party. Both games have a lot of potential, but hopefully they can transcend Façade in terms of broad appeal & replayability, and have a bigger impact on the community.

Expectation E.
Beyond Competence

I initially stumbled on a preview of the upcoming KILLZONE 3 multiplayer bots that called them "jerks" for the way they played. The bots in the prequel received extremely favorable reviews, but tended to use grenades a bit too often — and that came up in occasional comments. Now these issues have been addressed, what's left to comment on is the core behavior of the bots... and that's now being judged not by its minor flaws, but by its play style!

I thought this was a huge breakthrough at the time, but wondered if that was an isolated incident. Then while writing this editorial I stumbled on further reviews of the FORZA MOTORSPORT 3 AI that emphasize this trend. The racing AI in the game is very solid, but it plays a bit too aggressively 15% of the time and uses too much force to knock the player off the track. Again, this isn't a question of bugs, but play style.

My prediction for 2011 is that we'll see more of this. Game companies that are investing in AI will see their bots and opponents reach a certain level of competence that rules out common bugs, allowing journalists and gamers to focus their comments on the style of play and the personality behind the behaviors. Occasional games in the past have reached this level of competence and completeness in their AI, but I expect this to become a bit more systematic and widespread this year... Can we finally focus more time on the details of AI play and how they are perceived by the players?

Expectation F.
AI Directors as a Design Pattern

This year, ALIEN SWARM shipped and was made available freely by Valve as an example of its SteamWorks capabilities. Shortly after, the full SDK and code for the game was made available, and quickly the community discovered its 'hidden' AI Director and how to enable it. In a subsequent patch, this feature was enabled directly from the game menu, and this received a lot of attention from the press and gamers alike. The implementation of the AI Director in ALIEN SWARM isn't quite on par with LEFT 4 DEAD's due to the different mechanics in the game (no rescuing or reviving), but it's a fun addition to the game nonetheless.

Based on some pre-release marketing snippets for upcoming games, I expect AI Directors to become much more common in 2011. This is the case for DARK SPORE, among others, that is using an AI Director to guide its gameplay in a way that's fun for players. I hope to see many more AI Directors in the future to help move beyond fixed scripted experiences and help those linear stories fit player skills better.

Discussion 2 Comments

bweber on January 7th, 2011

I find your outlook based on CIG 2010 and AIIDE 2010 to be encouraging. It seems your grim view of game AI research going into CIG has turned more positive; as long as you are talking to the right researchers. My prediction for 2011: The rise of Procedural Reasoning System (PRS) frameworks in games.

alexjc on January 10th, 2011

Ben, Yeah, I think there's lots of positive stuff happening. It's a good management practice to focus on what you want and ignore the rest (until you can better deal with it), so that's going to be my focus too -- making sure academics who want & claim to be doing industry-directed stuff get better feedback. Also, I was emailed about my comments about the academic planning community. It seems there's quite some controversy in there, especially regarding HTN. But that's a blog post for another time! Alex

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!