What are the "issues" in Game AI [moderated thread]

From a developer perspective:

My most dominant concern? Resources required for algorithm/process X. If I can do a 10 ply alpha beta search on a chess board, and have it search the tree in a reasonable amount of time, that''s what I need. If I have to cut back to 8 ply to shorten the time it takes to search the tree, it''s better then having a move 20 minutes later then the player wants it...

It really doesn’t matter if I can make realistic AI, challenging AI, fun AI, if it can''t run within the game with the resources it''s meant to work with.

I find myself spending more time trying to increase performance, rather then doing research and making BETTER AI.

Really, game AI, like graphics, is not usually a matter of "what is the new cool technique", but rather "what technique will work with X resources".


If all that matters is what you get in the end, why go through life?
~Michael Sikora

From a player perspective:

The use of AI can depend greatly on the type of game that is produced. In general terms the target audience and genre figure greatly in the AI needs.

Personally it depends on what kind of mood I’m in as to what type of game I go for and thus what the AI is like.
For example if I just want some relaxing fun, running around blowing things up in a FPS, then the AI is less important than just bring on the bad guys for me to waste. However I still feel that they should have enough of an AI not to ‘toast’ each other needlessly, get around objects and to some extent gang-up and ambush me.
If I want a challenge however I’d go for a more strategy based game, either RTS or Turn Based Strategy. These I would expect to have a robust and complex AI, something that will prove a sufficient challenge. I also expect that in games where there are more than one opponent to face, that they aren’t allowed to just gang up on the human player in the old ‘All the AI’s together’. The other thing that you get a lot of in these types of games is the opponent getting free hand outs from the program, this I have a real problem with. Similarly with map / environmental knowledge. If I have to explore and conduct a hard slog to secure vital resources, I don’t see why my opponent shouldn’t have to.

Personal rant over.

I think, as I have said before, that it all boils down to time and money. Time pressure in releasing a product and tight budgets.




"Making it up! Why should I be making it up. Lifes bad enough as it is without wanting to invent more of it."

"developer" POV :
The unnerving dependency between AI and more elements of gameplay than one would like to imagine.
I have been "researching" for my MSc for a good year now, and the constant problem I run into is the fact that AI is not just a case of figuring out a few good algorithms to make your agents run around.

As soon as you start interacting with the environment and its agents, you need good interface design. which means, you need to know a bit about Human Computer Interaction.
It''s far from being a trivial problem to solve. From the "click to use" policy of HalfLife, to the simple generic orders of Unreal, to squad commanding in Operation Flashpoint. And let''s not get started about dialogue systems...

Another thing. If you think agent, you think "moving around", because let''s face it, movement is one of the first thing that let an agent look "intelligent". The problem is that movement, is highly dependent on the environment. That is, without the proper structures to describe an environment, how do you expect to move around ? You''d tell me, "hey, that''s not AI, that''s someone else''s job". Yeah, precisely, that''s the problem. Someone else, who has different goals than you when coding.
Think of the conflict between the artist who wants thousands of polys, and the coder who want to fit them in memory.
Well, the AI guy wants (OK, *I* would want) as many details about the environment as possible, while the coders don''t really care, they design the structure to hold polygons and textures.
If you dont understand what I am on about, think about it that way : the graphic coder will reduce a level to a set of polygons, organised as a nice BSP, with textures info. As an AI coder, I would rather have all those polygons organised as more "user friendly" sets of polygons, say rooms, corridors, walls, etc.
This is what a system like AAS (the system used by the Quake 3 bots) does, it recreates the level from the mass of polygons, and organise it in volumes where bots can move. It''s a step in the right direction, but I still think there could be more.

Another point that goes with this one is that with more animations, you can visually depict more behaviour. That is, if all my agent can do is walk around, it''s OK. But if my agent can crouch, go prone, walk, run, jump, etc... it starts getting pretty interesting. Imagine how cool it would be to look at a group of agents communicating with hand gestures. The gestures would of course only be there for the possible human player watching, but it would definitely be an expression of the intelligence of the agents.

Do I make any sense ?

Anyway, I am biased since those two points will probably end up as the main points of my report.

To summarize my ranting, AI is not the only thing an AI coder should be aware of. The structure describing the environment in which agents evolve has an enormous impact on the scope of behaviours that can be exhibited, and on the amount of information the agents can use to base their behaviours on.
Think of the step we made when we switched from grid based systems to polygon based systems (Eye of the Beholder vs. Morrowind, anyone ?)
Equally important is the interface that let us interact with those agents. Think of the difference between clicking on bottomless menus, and a speech recognition based interface.

I hope this contribute to the discussion. Any comment is welcome (if that''s OK with you timkin, and if you think I am on topic).

---

Sancte Isidore ora pro nobis !

developer:

AI is only as advanced or intelligent as it can be in a given environment -that is, it is limited by its ability to gain input and produce output. I like to write sims to study AI, and i find myself spending all of my time writting environments and their laws. once an environment is set up, there is the issue of input: how does the agent get input? alot of time it has sensors, which raises the the issue of what the sensor does, does it detect pixels or vectors, or does it detect entire objects? Next is output: what can be done, changed, or thought up?

After all of that is set up, you can begin to tinker with AI, and the ai is tailored to those things. the end result is a specific context, like the context of a game, and the "AI" can only function intelligently in that context. we are ultimatly trying to acheive ai that can function in the human context, and all ai does do that to a limited extent.

the player should be an important component of AI. too many games have static ai, that doesnt learn between games. AI needs to be able to learn from the player. players introduce elements that may be seemingly random, and elements that simply are beyond the AIs understanding. this is especially true in real time strategy AI. its not uncommon for players to exploit bugs in the ai to gain an advantage, in such areas, the only solution is to take a biological approach - and evolve, using genetics (or trial-and-error ).

evolution allows you to introduce random elements and see how they pan out -if the result is good, then the AI has learnt something new, perhaps previously beyond its grasp (and perhaps still). The implimentation of (practical) evolution can be tricky, but i would like to see the AI try some stuff for no apparent reason, and see how it works.

that in my opinion is the problem with most AI -the ai does not take risks. why cant it take risks? -it is a game after all, not a life and death struggle on the discovery channel. the game will get played again, perhaps the ai can request a rematch?


-evilcrap

Player perspective :
- EvilCrap brought this up above, but I''d like to reiterate. For almost any game, one you have beaten the AI you can beat it again continuously. The AI is no longer challenging to play against, and since you can''t tell it how badly it sux0rs it''s no longer fun (that last only partly in jest).
- I really dislike influence mapping in games (I think that''s what it is at least). From what I can tell this makes for very boring, very annoying AIs. Firstly it''s predictable and hence easily beaten as soon as one figures out where its valuation is imperfect. Secondly it seems to go hand-in-hand with total map knowledge which is aggravating (ie. you leave one spot and the computer will be there the next turn). Which brings me to my next point.
- Cheating. Almost every AI cheats in some way. The AI is usually playing under different (relaxed) rules which usually means the player has to exploit its stupidity in order to offset its artificial advantages. Discovering the stupidity of an AI makes it much less fun to play against.

Developer perspective :
- It''s hard to make an AI and not result in one of the above.

Nice thread. Though it would be of even more help if someone would summarise and regroup all the points in one line each (Timkin

player:

• Can I mess up the AI character by pushing it out of its sand-box, i.e. the domain for which it was designed?


• Does my interaction with the AI cast have any affect on it? Is it worth me spending time on/with/around it?

developer:

• How long would it take to explore and experiment with a particular interesting option, before it even makes the game?


• Is middleware viable?


• How to integrate all the academic research in practice?


• What are the options for making HUMAN-LIKE AI, and which is most applicable to my problem?

design:

• Can AI drive the design?


• Does the design get in the way of the AI?


• How can both be made to fit and match together?

Yes, interesting issues! Now what?

---

Artificial Intelligence Depot - Maybe it''s not all about graphics...

My brain puke for the hour:

As a designer, for fun factor:
* Is predictable AI part of the challenge of the game? (Run the monster into a corner to escape it)
* Will intelligent AI make the game too unpredictable for the player (consistent strategies don''t work). The idea is to help the player master the game, which is part of its draw.
* How much should the AI be dumbed up to make it seem more realistic? All shots shouldn''t be so accurate, etc.
* Settings for how AI reacts (easy, medium, hard), and have it actually BE easy, medium and hard!
* Should higher skill levels challenge the player to try new tactics or polish their existing ones?
* Don''t railroad only one possible way to effectively confront an AI opponent, if possible
* Does the AI remember or just work off from the existing state?

Player:
* Does the AI challenge me to try out a few tactics and stick with one that works? Do small variations of that help me confront other AI objects as well, given the interface constraints? Do variations in the environment challenge me to tweak my strategies given each situation?
* How much do I have to micromanage my AI units?
* Do my AI units do too much and take away from the amount of control I''d like to have?

Management:
* How modifiable is our AI system? If we want to add more or change it, will it take a lot of reorganization?
* Which AI algorithm is appropriate for what we want to do? (Fuzzy logic, neural networks, multi-level finite state machine, etc.)

quote:

---

Original post by alexjc
Though it would be of even more help if someone would summarise and regroup all the points in one line each (Timkin

---

That will be done, but for the moment I don''t want to unintentionally steer peoples thoughts by highlighting what has been said before. I''d rather people read it themselves and offer their own ideas.

So far, everything that has been written is great! There are certainly many points raised that I doubt I would have considered myself (which is why I am asking you folks.

Please keep the ideas coming, but also feel free to discuss what other people have said. Expand on their thoughts from your own perspective.

As to what it''s all for... I''ll let you know when it''s all said and done!

Thanks,

Timkin

It seems that people aren''t particularly interested in this thread. If anyone has ideas about how to better stimulate peoples minds and get them to commit to thoughts on ''e-paper'', please let me know.

Thanks,

Timkin

Maintain a list on the site and let people add items to it? I''m sure it would be useful.

Exactly, Argus is right. Reading through all those long posts to get some key points is very tedious, and I''m sure most get bored before the end. I posted 8th, and I confess not reading fully past the 4th message, and just scanning for keywords.

The time has come for the summarisation. Maintain a list right at the top of the thread (format it nicely in HTML), mention and emphasize the purpose of it, and tell people to add their comments as a new post. Also explain that a moderated thread is just as open as others... this might have scared people off a bit. Keep it concise! Short attention spans, rememb...

What was i doing again?

Alex

---

Artificial Intelligence Depot - Maybe it''s not all about graphics...

In Games:

Nowadays I think that developers put care in their AIs. I''m happy because it makes completely different games...

Starting from the beggining, and forgetting all the games that don''t have AI, I think that soccer games were the firsts trying to implement good AIs. However, Kick Off (well known thanks to its great playability) is very simple today. Sensible Soccer was a better game (with simple graphics) because it had a good AI and, due to that, a great playability. Fifa Series are not liked by soccer-videogame players because it is too easy and AI is very stupid. Look at ISS Pro and, though its graphics are not so outstanding, the playability is great.

So, what do I mean? In EVERY game that requires AI (at your side or as an opponent) the people is expecting better bots.

The same happened with the first bots for Quake, and what people did later (I think it was Mr. Elusive who introduced fuzzy logic in Quake bots, but I''m not sure). Now, Quake bots play far better than in Quake I.

Another good example of well invested funds in AI is GTA3. The drivers are really good, they act as real humans and take real good decisions when driving (escaping over all). Another one: black&white, the people acted REALLY well, but if you played the game you should have noticed that the opponents (AI gods) where GREAT.

Design: how to get this, how to get a GOOD AI?

This is the real problem. As designers I think that our goal is:

- Get the AI difficult to beat
- The AI can''t play always the same way (so the AI must learn or use random factors)
- The AI must not have cheats that can lead the player to exloit them (and lose fun and interest)
- The AI must adapt to any way of playing allowed in the game
- The AI shouldn''t have to cheat to get these goals

So, we can see many circumstances where this is not aquired: urban terror bots, cooperative games, heroes of might and magic 3 (I do think that homm3 AI cheats all the time)...

For me, the problem is "how to model" the AI. In many cases a combination of techniques is used, and balancing and selecting them is a difficult task. For example, programmers have been studying chess for years and modern approach is far different for that used in the beggining. Maybe chess is not a good example but for me it shows how an unique problem can be extended for years of investigation.

For this reason, getting a good AI system the first time you model and code it should be difficult...

I think that, once you master AI techniques, you can start aquiring some "intuition" that lets you think about a problem and a suitable way to perhaps solve it... and it can not be the best way...

Even if you master many techniques you still have to code some algorithms (asigning weighs, heuristics, estimating states importance, designing a neural network, modeling your genetic algorithm...). Where do you get all this? In most cases you model them because you suppose which are the important things and which ones are not important but, still, you should do many tests and demos and check how the AI works in every case.

So, I think that AIs are like big art pieces (little in my case), you work on them, at one point you are satisfied or don''t have more time and you end with a little system that can play as a human would. Some are really good, some are impressive, some are fun and surprise you with well made responses to things that you think that wouldn''t be noticed by the AI...

And all of them are different...

Of course, from a scientific point of view you can think that a problem can be studied and reduced... well, in some way it can but in that case maybe you are forgetting to compare some technique... and it takes too long. I don''t know but I think that designing (modelling) AIs is an inexact science. And maybe that is what makes it fun, you always end up looking at it for hours and wondering why your AI does or doesn''t do that thing...

In the other hand, some techniques areuseful for pathfinding and so, but, though they are "AI techniques", doesn''t make an issue since they are already studied and work well. For example, you may use AI techniques to move people in a graphic adventure, but this is not what I am refering in the post and I think that everybody here is talking about AIs acting as human players, or as intelligent beings inside a game.

Anybody here plays bot fights? Someone can give me links or something? We could make an arena and a scripting language anyway (if anyone is interested I''ll post this in another forum).


JJpRiVaTe
Web: Private Zone

As a player:
The All-I-ever-though-about-was-killing-you-and-everything-that-looks-like-you behaviour from, say a bot is the worst thing I know (by that I mean bots that rush forwards and shoot and shoot and then repeats from rushing forwards)
Seeing monsters or NPC''s or whatever cooperate to achive a shared goal impresses me alot.
Seeing monsters or whatever hiding or doing other things that seems to need a little bit of intelligence impresses me.
Being able to hide from a monster by standing still in the shadows would make me cry of joy.

I guess that''s it.

Biggest problem for me is creating an AI that is all things to all players. As we''ve seen from some of the posts, players all expect different things from their AI. In addition, different players have different skill levels. Trying to cater to all at the same time is a nightmare.

Bob

My main concern and area of interest is in creating non-player characters with believable behaviour (specifically for RPG-like games). It surprises me how little advancement has been made in this area, given that a hybrid script and reactive system should be able to create something very authentic.

[ MSVC Fixes | STL | SDL | Game AI | Sockets | C++ Faq Lite | Boost | Asking Questions | Organising code files ]

I have never written any AI for real time games other than for pong. The AI that I have quite a bit of experience with are things like chess, checkers, etc. That area of AI makes use of mostly brute force methods. Basically the AI is the program, so it''s different than a real time 3D shooting game.

As a user, the thing that I find to be the biggest drawback is that there is not a realistic feel to the game play against artificial opponents. I''ll give some examples. In Quake 3 Arena I play against the bots on the hardest setting and I am able to run straight to the flag, grab it, and capture it most of the time. Not always though. The times that I don''t make it are usually because I get shot while flying through the air from across the map, which is also a little unrealistic. In Counter-strike (Half-life mod) I play against bots, and it''s very easy to dominate them even when they''re on the hardest setting. I can play 1 on 10, and because they are so predictable I can usually take them all out. Here again, when I fail it''s usually because one bot glocked me from across the map (glock is the biggest POS gun in CS). This is fine if I want to practice my sniping or something, but for "fun" it gets very old very fast.

As a developer, I (like most I would guess) don''t really have a solution. I want my AI to be more realistic and perform like a human and make mistakes sometimes. But then you get this fake "mistake" crap where the bots are running the wrong way into a wall while they should be going for the flag or defusing the bomb. I think that current AI seems very "hacked together", where they first develop a bot that plays almost perfect and never misses, so they dumb it up a little, and then it has this unnatural feel where it plays incredible but all of the sudden goes nuts and does something stupid, or it wanders aimlessly then kills you in one shot from across the map.

Even in chess the computer doesn''t play like a human. The computer sees quite a few moves ahead so it can thrash 99% of us humans, but it has a limit and can''t see THAT far ahead and see the long term effects of it''s decisions. That is how top chess players are able to beat chess programs. The computer can see maybe 14 moves ahead, but the chess grandmaster knows that a particular feature of his computer opponent''s position will come back to haunt him 30 moves down the road. On top if this, the computer plays moves that don''t make sense a lot of times. The computer will make what we call "shuffling moves" where it kind of "does nothing". This approach works well against weaker players, since it''s only a matter of time before a weak player makes a weak move. Against a computer it only takes one slightly weak move for the computer to hammer you. The computer chess program has no concept of a plan. It makes move A instead of move B because there is an array of square values that says that the center squares are "worth" more than the border squares.

Having said all of that about computer chess, here is what I think. The large portion of the time in a real time game is spent on non-AI things. In computer chess, 99.9% of the time is spent on the AI, and maybe a very small fraction of the time is spent on generating output or something. Computer chess still has many holes and is still very unnatural when compared with playing a human opponent. The fact that computers beat humans most of the time doesn''t mean the AI is good. IMO it''s quire poor actually. It''s nothing more than a pretty basic evaluation function and some standard search techniques. There are a few interesting pruning techniques such as null-move pruning where the program decides not to search an entire branch of the tree without even searching it, but even still that''s not very advanced. They aren''t really smart, they''re just very fast. In other words, it still hasn''t been done in computer chess where a program plays like a human. A program cannot give you a very in depth analysis of a position. It can give you a score and some moves, but it can''t output, "The reason that Nxf6 is a bad move is because it allows black to develop a strong attack on the queenside." About the best "analysis" I''ve seen comes in the form of general statments such as, "After exd4 white''s pawn structure is slightly weakened." The program doesn''t "know" that, it just got a worse numerical score because it considers an isolated pawn to be bad always. This isn''t always the case. Sometimes it can be used as an advantage. The computer cannot decide whether it''s good in this case or not. It just assumes it''s bad because it''s bad *most* of the time. It uses a lot of these general rules, and sometimes these rules make it look very dumb.

Since in a chess program (which is 99.9% AI) the computer can''t achieve a good understanding of a chess position, I certainly don''t think that in a real time game where a small fraction of the time is devoted to AI that the AI has much chance of being realistic. The first thing obviosly is that the program has only a small fraction of the time to do AI stuff when compared with chess. Secondly, in a real time game the environment is constantly changing. By the time the computer has had time to analyze anything, the environment isn''t the same, so the computer''s decision to "go left" might not be a good decision anymore.

I think real time AI is a very difficult subject. My personal opinion is that it''s quite a ways off too. I don''t think we''re just waiting for the "next big thing" before we have realistic AI. I think we''re going to be waiting for "the next big thing" after maybe half a dozen or so "next big thing"''s have been discovered.

Russell

Again I want to thank everyone that has taken the time to respond to this thread. Your responses show that you have given thought to my question and spent some time considering your responses. I appreciate that.

I'd like to raise a couple of further questions, in response to some of the points made. Feel free to discuss them further.

Guardian_Light mentioned the problem of 'efficiency' and 'performance' of AI in terms of the resources available. Clearly one of the problems in developing AI for specific game platforms (XBox, PS2, etc) is the limited resources.

Do people feel that a well designed and developed AI middleware package would solve this problem by:

a) providing a clear understanding of the resources required to implement any given technique;
b) encouraging the development of AI dedicated hardware; and,
c) give more people access to the AI techniques that they don't have time to develop properly themselves?

Several people raised the issue of 'reality' in AI. Whether an AI 'looks' and 'feels' real to play against. Let's face it, most if not all AI's just don't feel real (IMHO).

Would you, as a player, believe an AI to be more real if the mistakes it made appeared to be flaws in its strategy or tactics, rather than ad hoc dumbness (like poor aim) or sending exceptionally weak squads against you one after the other.

What do you, as a player, believe 'acting like a human opponent' means?

Another very interesting point raised was the issue of adaptability of the AI. Does the AI encourage adaptability in the player and does the AI itself adapt to the players altered tactics and strategies. For the most part, the player is given different levels to complete, usually of increasing difficulty, that encourage adaptation or refinement of strategies and tactics.

Do you believe that imbuing an AI with the ability to adapt to the straetegies and tactics of its particular opponents - with the aim of defeating those strategies and tactics - would make for an interesting and challenging AI, or do you believe that it would be too hard for most players to overcome the AI and hence not enjoyable?

Given that the AI could be made to adapt to the player - through modelling of player behaviours and planning of effective responses - how would you go about weakening the AI without making it seem like an obvious hack?

Finally, Russell and several others raised a very good point about whether the design of the game was driven by the type or needs of the AI, or whether the AI was developed after the gameplay was designed (i.e., to meet the needs and vision of the design team).

Do you feel that games would be better served by taking a more AI-centric approach to design and gameplay, just as games like Doom and Quake once to a very graphics-centric approach to attract players?

Again, thanks for all of the responses... and feel free to raise new issues and to discuss them.

Thanks,

Timkin

[edited by - Timkin on June 30, 2002 10:51:28 PM]

i think ppl must set AI above graphics now. After Doom & Heretic, there has been a tremendous battle about the graphics. Now this should be started for AI too. Then we get the best AI :D

quote:

---

Original post by Timkin
a) providing a clear understanding of the resources required to implement any given technique;
b) encouraging the development of AI dedicated hardware; and,
c) give more people access to the AI techniques that they don''t have time to develop properly themselves?

---

a) IMO, the greatest advantage of a middleware is the ability to try a possible solution quickly, so it would probably provide a clearer understanding of the resource requirements. Right now, I assume a lot of developpers must choose a technique at the beginning, and don''t have time to try something else if they find out that they are wrong.

b) I don''t think so, because a general-purpose CPU will probably be favored (in games, at least). But then, I have 0 professional industry experience, so what do I know about it?

c) Definitely, especially with smaller teams or budget titles.

quote:

---

Do you believe that imbuing an AI with the ability to adapt to the straetegies and tactics of its particular opponents - with the aim of defeating those strategies and tactics - would make for an interesting and challenging AI, or do you believe that it would be too hard for most players to overcome the AI and hence not enjoyable?

---

My father used to play StarCraft all the time, only against computer opponents, because he could figure out the pattern, and play StarCraft as he played a game of Solitaire. When you''re a good player, it''s great to have a changing opponent, but when you''re not, sometimes the only way to beat a particular challenge is to go through it many, many times and figure out exactly what moves must be done at each step, to compensate for a lack of skills.

quote:

---

Given that the AI could be made to adapt to the player - through modelling of player behaviours and planning of effective responses - how would you go about weakening the AI without making it seem like an obvious hack?

---

There is a great way to do this in Chess: the computer has either a limited horizon, or it is forced to make a move after a short period of time. If AI was designed to assign scores to particular moves, based on its analysis of the situation, limiting its thinking time could be an option (it would also allow ''soft-core'' gamers to improve their framerate by limiting the AI''s ressources).

Not much insight here, but I had to write a response to ask you: You''re working on AI middleware, Timkin? When/what/why?

Cédric