Slashdot Mirror
Artificial Intelligence for Computer Games
Funge's background includes both academic AI research and commercial development of game AI technology. This has allowed him to write a refreshingly practical book for the game AI programmer which will also expand the reader's knowledge of AI. He presents advanced AI research in a way that is meaningful to the working game AI programmer. Non-player characters (NPCs) are the focus of this book, although it touches upon techniques applicable to other kinds of AI. Funge begins with a simple NPC architecture, then goes on to consider how they act in their world, perceive and react to their surroundings, remember their past experiences, plan their actions, and learn from the past to improve their future behavior. In addition, Funge hopes his book will contribute to a "common framework and terminology" to promote better communication between practitioners interested in game AI, leading to better interoperability for their software. (Please note that John Funge is a friend and former coworker of mine. I was pleased to accept John's invitation to review his book.)
The field of Artificial Intelligence has been actively studied since the 1950s. In that half century many useful techniques have been developed and applied to a broad range of scholarly and commercial applications -- most quite serious and sometimes a bit dry. In contrast, today the most economically significant application of AI is in computer games. This commercial application motivates today's students to study AI and drives a good deal of academic AI research. Modern games have incredible graphics and their animation technology is becoming very sophisticated. As graphic animation increasingly becomes a solved problem, more and more attention is being paid to game AI. It seems likely that the next few years will see a tremendous investment in game AI technology leading to significant improvements in the state of the art.
As I read Funge's book I was struck by how oriented it was to the interests of AI programmers working on commercial games. Certainly the discussion focused on the practical rather than the theoretical. (There are many asides, footnotes and citations of the academic literature for those with an interest in pursuing the theory.) More concretely, the text is peppered with fragments of C++ code. A working programmer who visits the academic literature is often faced with the daunting task of converting prose, equations or breezy pseudo-code into something suitable for compilation. If a reader of this book does not follow a bit of the discussion, a glance at the nearby C++ code listing will usually set things straight. I have it on good authority that functioning source code for the examples in the book will appear on the www.ai4games.org website "soon."
The book is divided into seven chapters (Introduction, Acting, Perceiving, Reacting, Remembering, Searching, and Learning) plus a Preface, two appendices, an extensive Bibliography and an Index. The chapter on "Acting" introduces the simple game of tag used as an example throughout the book. It further sets the stage by describing the principal components of the game engine and the AI system. The third chapter, "Perceiving," introduces percepts -- the formal framework used to encapsulate and manipulate an NPC's awareness of its world. In many games a key concept is filtering out information which is available in the game state but should not be "known" by the NPC. Chapter 4 describes reactive controllers. Funge uses a very strict definition of reactive -- informally, it means a non-deliberative controller, but in this book the term is used to mean strictly stateless. This distinction has a practical consequence since a stateless controller can be shared among multiple NPCs. (Yet I wondered how important this was in practice. That point was not explored in any depth, and a "slightly stateful" reactive controller can be very useful.) The chapter on "Remembering" introduces memory percepts, mental state, beliefs and communication between NPCs. The sixth chapter covers "Searching" -- through trees of possible future actions, often referred to as planning. The extensive treatment of search includes both examining the host of options that are available to an NPC at each juncture, as well as reasoning about the interaction of one NPC's behavior with another, known as adversarial search. The final chapter covers "Learning." It looks at both offline learning (which happens before the game is shipped) and online learning (happening during gameplay). The first is merely an aid to game development, the latter promises NPC that can adjust to the player's skill and style of play. Online learning present many more technical challenges. In fact, my first impression on reading this section that it was less practical than the rest of the book because of the difficulties of online learning. However, from the description of this GDC 2005 lecture, it appears that Funge and his colleagues have made significant progress in this area.
I recommend Artificial Intelligence for Computer Games: An Introduction to commercial game AI programmers, as well as other game programmers and designers who wish to learn more about this area. Because of its sound academic underpinning, the book will also be of interest to students of artificial intelligence and to professionals in related areas such as agent-based simulation and training.
Reynolds is a Senior Research Scientist in the R&D group of Sony Computer Entertainment America. His interests center on modeling behavior of autonomous characters, particularly steering behaviors for agile life-like motion through their worlds. See his page on Game Research and Technology. You can purchase Artificial Intelligence for Computer Games: An Introduction from bn.com. Slashdot welcomes readers' book reviews -- to see your own review here, read the book review guidelines, then visit the submission page.
What I need is a game with AI that can evaluate my game play and tell me how to play better against my opponents, kind of like where you view your opposing team's old games to learn their patterns and weaknesses, only give me feedback in real time while I'm playing.
The NSA: The only part of the US government that actually listens.
I have yet to find suitable AI for that game. To me that type of AI (dealing with imperfect information) is the holy grail of game AI. This includes a vast family of such games (poker comes to mind), where over time, information about the game state is disclosed. I once found a small Stratego game on the net that played an unbelievably good game. After saving and restoring a game a few times and losing (whereup it shows you its pieces) it had actually cheated real-time and would move its pieces to match yours. Thus the small executable! :)
Hedley
Buy an old C64 or Amiga copy still in the box if you can. Seriously, I mean it. It comes with 2 manuals.
Book 1 has a short story and some player info, and Book 2 describes step-by-step exactly how they developed the AI for the game. Going to cons and watching successful players play, getting them to give out their strategies, and then translating those ideas into code. As a bonus, they describe the exact formulas the Ogre uses to determine its move, targeting sequence and deployment of arms.
It's brilliant, informative, and well worth the price of the game alone. Highly recommended reading if you're into game AI.
Weaselmancer
rediculous.
And I don't just mean "it's not technically impressive". I mean, considering the purpose of in-game AI, it's *bad* AI. Good AI generally simulates more complicated human interactions. Good AI can be tricked or distracted, and can learn so it's not so easily tricked the next time. I really like the idea of AI that will adjust to the player's skill level to always provide gameplay that is exciting and challenging, yet beatable.
In other words, I believe "good" AI in a game is not defined by being hard to beat, but by being fun to play against.
The next "killer app" for MMOGs is advanced, learning AI. Right now games are trying to cover up the simplistic behavior of NPCs by creating complex scripts around them.
Example:
- Between 100%-75% health, Dragon will fight as normal.
- At 75% health, the dragon will breath fire in attempt to kill as many players nearby, fly over to the west part of the chamber.
- Between 75%-50% health, Dragon will fight as normal and start using its tail.
- At 50% health, Dragon will fly to the east part of the chamber, breath fire onto the players as they run from the west part of the room to the east.
So on and so forth. The problem is that humans easily can see paterns like this. This "event driven" behavior only works when players are "surprised" and becomes a serious liability when players discover the pattern. If the pattern is "discovered", players will scatter around 75% to avoid the fire. At 50% they will run to the eastern part of the room before the dragon gets there to avoid it breathing fire onto the western half.
To avoid some of this predictability, some monsters appear to have "randomized behavior". A monster has 5 different "actions" where a programer weights the choices and generates a random number. This makes the monster appear to have some tactics trying different attacks but just as much as it succeeds in throwing the player off they will often randomly chose the poor action.
I believe advanced AI techniques like nueral nets will be the next "killer app" for MMOGs. Learning AI is not impractical for a single player stand alone game but it is not as "exciting" nor do single player system have enough computing power and "experience" to really put a nueral net through its paces.
The Dragon in the example starts out like the players in that neither side knows exactly how to win. Reguardless of the outcome both the sever/Dragon and the players should learn something from the encounter. Have enough players run against The Dragon and it might start to learn things like "fire seems to be more effective against melee". When it sees a raid comprised of mostly melee and very few casters it choses its fire attack far more than its melee. This is a far better option than "randomizing attacks" or scripting their behavior. The Dragon is now actually using tactics and reacting to the players in a psuedo-intelegent manner.
If we really want to go far fetched, it would be great if each server instance of The Dragon "learned" on its own and developed personality and behavior unique unto itself. One server's Dragon might like to fly around compared to another that likes to walk when moving around. Of course one of the tricks is keeping the game engaging. No one wants to fight The Dragon if they know it will beat them 9/10 times.
Some NPCs should be designed simplistically because that is their nature. Some NPCs are highly intelegent and should act occordingly. I await the day when you can do true tactical attacks against the computer instead of having to resort to a scirpted monster or just filling the other side with other human players.
Some of you might be aware that the PC/Mac/Linux Game Neverwinter Nights includes a toolset with a C-like scripting language that allows users to code the behavior of characters in a game---not just for combat, but generic interactions as well.
BioWare, the developers of the game, are known for the imaginative story lines in their Star Wars: Knights of the Old Republic and Baldur's Gate series. However, by their own admission, they never have as much time as they want to work on creature AI. In Neverwinter Nights, this shortage of time resulted in a number of unfortunate situations during game play. For example, friendly characters would waste powerful spells on pitifully weak enemies; or they would continually attempt to cast spells in close hand-to-hand combat, not realizing that this gives the close-by enemy countless opportunities to tear them into pieces, and that pulling out that dagger in their backpack might be a better idea. Especially sad were near-death enemies who would try to heal themselves with woefully inadequate healing spells (in RPG talk, down 80 hit points and casting cure minor wounds).
Luckily, the toolset allowed a number of us to code improvements to NPC behavior. I was one of them, starting the Henchman Inventory and Battle AI project, now lead by Tony K. The focus of our project was immediate improvement of game play. An even more impressive community is the Memetic AI group. These folks are putting together a full package of complex behaviors for an entire world, from peasant farmers to fearsome dragons. Impressive stuff.
I'm totally with you on the ability of people to recognise patterned behaviour, typically far faster than a game designer might suspect.
:)
But people can also discover "emergant" patterns that aren't necessarily explicitly programmed in.
I remember playing Sargon III Chess on my C-64. I accidentally discovered that the AI couldn't see - I'm no chess geek, so I'm sure there's an official term - "indirect" attacks. Rather than move piece A to sqaure Q ro threaten enemy piece X, I'd move some other piece B onto the line of attack that I wanted to make A->X, blocking the attack. Then piece A would be moved into attack position on Q, and piece B moved out of the way, unblocking the A->X attack.
The AI seemed to be able to predict that a straight move to Q by A would threaten X, and it would be very good at countering those moves. But attacks from a third piece by moving some other piece out of the attack line were invisible to it.
Once discovered, this lead to strategies that involved setting up elabourate attacks that hinged upon "reveals". It'd drive the AI nuts. Sadly, actual humans do not suffer from this blind spot and ol' Sargon did not improve my RL chess playing ability one bit.
Here's another example of a different kind:
One summer, a group of my friends played a TON of the original Battletech board game aginst each other. We'd start after supper and go to the wee hours of the morning, day after day after day.
In so doing, we developed a particularly effective strategy. We'd have a 4-lance company. The first lance was composed of stripped-down lightweights equipped with maximum jump jet capacity and a single weapon - a flamer. The second lance was of superheavy, very low-mobility, weakly-armoured, long-range rocket artillery units. The third was ultra-heavy, low mobility, heavily armoured massive close-in-damage units, and the fourth was the reserve unit of heavy cannon equipped hovercraft.
As is typical for wargames, the faster you move, the harder you are to hit. There was a further negative modifier if the 'mech was jumping. Our lightweights, if they jumped full distance every turn, accumulated so many negative to-hit modifiers that they were unhittable. They would fan out over the game board, spotting the enemy and setting fire to terrain - which in the game rules, happened 100% with the use of the flamer - and which caused vision blocking due to smoke, plus there was a chance for the fire to spread to adjacent hexes.
The lightweights could also spot for the indirect fire lance with minimal penalties. The indirect fire lance would never move; it would just fire salvo after salvo of long range missiles. The hit rate wasn't great and the distribution of LRM fire tends to spread damage easily, but enough would hit as to ablate off some enemy armour - and the psychological effect of taking damage from an unseen source without the ability to retaliate... it was maddening.
Meanwhile, the heavy, close-in units would slowly advance up to intercept positions. Thanks to the madly-hopping lightweights and the smoke, we'd know where the enemy was but the enemy wouldn't know where we were.
The enemy would thus blunder up against the close-in units, which did monster amounts of damage with a high hit probability (the enemy unit was often moven slowly, due to the smoke, and the close-in unit would be stationary). It was not unusual to destroy an enemy unit in a single turn.
If things got sticky for whatever reason, the hovercraft would race in from the flank/rear and could disrupt the most cleverly planned counterattacks.
With all the practice we got, these tactics became drills - they could very easily have been scripted.
We put this to the test at a wargame convention, and we slaughtered everybody, without losing a single 'mech in any battle. Towards the end, the organizers were matching us upwards of 3 to 1; we just could not be beat.
Needless to say, we were not invited back.
DG
Want to learn about race cars? Read my Book