Perhaps you would be interested in embodied animats, where embodiment referes to the seperation of the brain and the body. The basic idea behind this approach is that an animat is more reactive to the environment it is put in because it uses sensory. I picked this information up from a book written by a gentleman named Alex Champandard (I hope I spelled that name correctly) He has designed a little system that works from the quake 2 engine called fear, I belive fear is on source forge but don't quote me on that... The name of the book is AI Game Development, and it discusses animats with the use of Neural networks, decision trees, genetic classifiers and much more...heh, not trying to advertise a book or anything, but if you are interested in some new approaches, this is a good thing to look into. Just my 2 cents

Quote: Original post by Nox587
I picked this information up from a book written by a gentleman named Alex Champandard

Hehe.. Alex has been a long time member of GameDev... although I haven't seen him around for a while. Having been a reviewer for this book in its production, I can certainly recommend it as good reading! ;)

Quote: Original post by owl
As other people say, a good game AI to me would be one that can play an learn to play better using exactly the same information the human player is given.

In the case of FPS shooting, the AI should have to do pattern recognition to target the enemies, at least.

First, Yay gamedev staff for fixing the servers. I can (finally) add my two cents.


That type of AI that continues to learn and do better is wonderful at an AI convention as a polite conference goer. ... But I feel they have no part in a game.


A quickie slash-and-burn player will face an easy AI at the end, since the AI won't have had time to train and improve.

A casual player will face a difficult AI at the end, since they tend to spend more average time playing yet don't spend time improving their own twitch-respose to the game.

A player who goes through from beginning to just before the end, then backtracks to the beginning, then goes looking for every concievable back door and flaw, is going to face an AI that is impossible to beat.

That's why I drew the distinction I said before... The ideal AI to have in a competative game is certainly not the ever-improving, self-training, and eventually perfecting style of AI.

The ideal AI in a game is one that provides a realistic, game-appropriate challenge. Whatever that is.

frob.

Quote: Original post by Anonymous Poster
The ideal AI in a game is one that provides a realistic, game-appropriate challenge. Whatever that is.

Well, that's what I was talking about. :) Realistic in the sense of playing against someone/something that is able to give fight without cheating. That's game-appropiate challenge to me :)

Quote: Original post by Anonymous Poster

Quote: Original post by owl
As other people say, a good game AI to me would be one that can play an learn to play better using exactly the same information the human player is given.

In the case of FPS shooting, the AI should have to do pattern recognition to target the enemies, at least.

First, Yay gamedev staff for fixing the servers. I can (finally) add my two cents.


That type of AI that continues to learn and do better is wonderful at an AI convention as a polite conference goer. ... But I feel they have no part in a game.


A quickie slash-and-burn player will face an easy AI at the end, since the AI won't have had time to train and improve.

A casual player will face a difficult AI at the end, since they tend to spend more average time playing yet don't spend time improving their own twitch-respose to the game.

A player who goes through from beginning to just before the end, then backtracks to the beginning, then goes looking for every concievable back door and flaw, is going to face an AI that is impossible to beat.

That's why I drew the distinction I said before... The ideal AI to have in a competative game is certainly not the ever-improving, self-training, and eventually perfecting style of AI.

The ideal AI in a game is one that provides a realistic, game-appropriate challenge. Whatever that is.

frob.

Interesting, I remember in the game Test Drive 4 the AI controlled cars would get faster if the player got faster, even when you chose a car that should be way out of their league and cut your lap times in half. That thing was crazy, it had to be cheating.

in my opinion GOOD artificial intelligence should havea degree of artifical stupidity

like if you attach a good AI to a FPS player model, it would know it's objectives but sometimes get lost, sometimes not know where a sniper is shooting from without continous ducking-and-looking to check, sometimes missing shots, panicking and running for powerups when health is low etc.

Artificial Intelligence is not intelligence at all. In fact it is just a lot of calculations and things like that.



Parts of the most advanced AI - according to Vexorian (which is most of the times wrong btw)

* Artificial Gameplay Well the AI should be able to play and win games correctly. Some older games had problems with AI in which sometimes it just failed, for example warcraft 2 if the map's ground was somehow hazardous the AI could get stuck there and just build units like crazy without using them...

* Artificial Learning The AI is able to learn tactics from the player, not only how to deal with them but even to return those tactics to the player or other players if necessarily.

* Artificial Stupidity The objective is to make a bot that behaves like a player would, no human is perfect so you'd have to simulate errors, and lack of information. So it would also have to pretend to gather information (in an rts it would need to scout for your base, most rts AIs already know where are you.

* Artificial personality If you have 2 or more bots in the same game, and you want realistic "intelligence" then they should behave in different ways. And not only for games with more than 1 bot but games where you could choose (or the game chooses randomly) how the AI opponent would behave, and battling 2 different AI opponents must be different, not only in difficulty but must give the sense of "personality" . Take Chessmaster's AI for example you can battle the Chessmaster but it also has some of the Legendary Chess players simulated in there.

Well im a university student (City Uni, London) and was one of the first students to do the Games design modules there. We were taught that good AI will "let you win with one bullet left in the clip"

As far as how to do this, well from my studies there was a very interesting comparison i noticed: The major drawback of OO Programming is the trade-off between too few classes that don't represent the world well too many classes that resemble the real world too closely.

too few clases:
take all ingame knowledge, eg. exact position of opponent in an FPS, and fire weapon at that point. repeat until no more enemies.

too many classes:
If we are to make a game AI seem to be Human (whic implies a LEVEL of intelligence) we would have to try and make an "Artificial Brain" that would take input from the world, but only for what the human would be able to see, eg a map and field of sight, and mesages from commander/ "friends". Then it would do some maths on it, check lookup tables for what "the real" human would do with this (suggestion) and do that. [insert millions of additional behaviours here]

but thats my thought. (and i'm very unexperienced)

Murtaza
Student (probably evidently)

Quote: Original post by Anonymous Poster


A quickie slash-and-burn player will face an easy AI at the end, since the AI won't have had time to train and improve.

A casual player will face a difficult AI at the end, since they tend to spend more average time playing yet don't spend time improving their own twitch-respose to the game.

A player who goes through from beginning to just before the end, then backtracks to the beginning, then goes looking for every concievable back door and flaw, is going to face an AI that is impossible to beat.

frob.

And as far as this goes, I think that AI should adapt, ofcourse, but it should have different "mindests" or, if you will, states. so we get an FSM.

eg.

Quote: Original post by Anonymous Poster

A quickie slash-and-burn player will face an easy AI at the end, since the AI won't have had time to train and improve.
frob.

The AI's state changes to maybe "sudden defence" and shoots anything that moves and runs around like its on drugs, and so the human is faced with a challenge.

Quote: Original post by Anonymous Poster
A casual player will face a difficult AI at the end, since they tend to spend more average time playing yet don't spend time improving their own twitch-respose to the game.
frob.

for this the AI is mirroring the player, so if the player attacks the AI will go to state "defend" and, well, defend. similarly if the player retreats after an attack the AI can go to counter-attack

Quote: Original post by Anonymous Poster
A player who goes through from beginning to just before the end, then backtracks to the beginning, then goes looking for every concievable back door and flaw, is going to face an AI that is impossible to beat.
frob.

Games aren't really made to do this, and so the player should load from a point where he (or she) can explore more, designers and programmers have enough to do :-)


I think that although they have states they can learn things like accuracy, eg if they player sucks set it's own to low and vice versa. BUT i don't know if using how the player plays is a good idea, since a normal FPS player would camp if they saw the enemy rushing all the time, so this could be the "Sudden defence" state. Or it could change the state definitions SLIGHTLY as things hapen, rather than create entire behaviours mimicking the player...


again, my unexperienced, student opinion...

Murtaza
Student :-)

[Edited by - m_khaku on November 22, 2005 9:08:25 AM]

Quote: Original post by Anonymous Poster

Quote: Original post by owl
As other people say, a good game AI to me would be one that can play an learn to play better using exactly the same information the human player is given.

In the case of FPS shooting, the AI should have to do pattern recognition to target the enemies, at least.

First, Yay gamedev staff for fixing the servers. I can (finally) add my two cents.


That type of AI that continues to learn and do better is wonderful at an AI convention as a polite conference goer. ... But I feel they have no part in a game.


A quickie slash-and-burn player will face an easy AI at the end, since the AI won't have had time to train and improve.

A casual player will face a difficult AI at the end, since they tend to spend more average time playing yet don't spend time improving their own twitch-respose to the game.

A player who goes through from beginning to just before the end, then backtracks to the beginning, then goes looking for every concievable back door and flaw, is going to face an AI that is impossible to beat.

That's why I drew the distinction I said before... The ideal AI to have in a competative game is certainly not the ever-improving, self-training, and eventually perfecting style of AI.

The ideal AI in a game is one that provides a realistic, game-appropriate challenge. Whatever that is.

frob.

FORGET FPS... Comparing an AI on fps is insane for a couple of reasons....

Firstly FPS requires non-complex calculations at high speed where humans are good at complex calcs at low speed..

Look to something like RTS AI, and stop thinking single player, think ORTS, think an AI that learns in RTS but that plays against many opponents, human and computer.

Thus the AI needs 2 make decisions based on,,,, intution/random, decisions like shall i rush and risk the game... Decisions like, so far i have seen 2 tanks, shall i build tank countering units and then he only has 2 tanks and get owned by infantry.

TO TEST AI systems you to get a very strong/state of the art u could say ai you need constraints!

An RTS has constraints, Game Speed/Game Mechanics/Limited Info about enemy player and so on.
-------------------------------------------------------------------------------
------ Q3 has a simple learning system, i know cause i have seen it learn..

My dad always camps on a certain map close to the BFG (big gun) and then owns the bots if they enter the room... So after playing the map 3-4 times, the bots just dont enter the room anymore, thus in turn prompting the human player to learn.

However i agree with Anonymous poster, having a single player fps ai is ridiculous! It means if i let my brother try to play a little then the entire AI would be lowered potentialy meaning i could own the final boss in 1 hit?