6 replies to this topic

[akramsulaiman@hotmail.com]

I am a university student doing a project on GENETIC ALGORITHM, FUZZY LOGIC and NEURAL NETWORKS and I would be grateful if you could give me some information on the challenges and problems that have hindered the use of the above three game programming techniques in game development.
Awaiting an early reply.
Thank You,

Akram.

[dave_]

Don't use so many capitals.
You'd also get better answers if you do some preliminary research on the area and ask more specific questions.

Do you think your tutor/lecturer/professor would be happy knowing you are posting this question on a forum like this?

You should look here first

[Nils Pipenbrinck]

The point of all these techniques is, that they have their area in applications where a system (e.g. the AI of a game) has to adapt to a "world" it does understand. If the world is relative simple, these techniques can - to some extend - lern the rules of a game.

BUT:

To the programmers of an AI, the rules of the game are very well known, so there is no need to train a network or genetic algorithm to understand things the programmer can put into simple tables or "if then else"-constructs. You, as the AI-guy know the rules in the first place. You can, with simple numeric techiques such as covariance analysis and statistics do much better than a neuronal network or genetic algorithm.

After all: A neuronal network is - in essence - nothing but a table of numbers. You can get this numbers if you have no idea about the world via a NN, but you get better weights if you know the worlds. In most cases you can calculate a global maximum of AI performance with algebra and statistics.

Some "AI lerns from the player" mechanics can always be integrated without using NN or genetics algorithms. The famous Soul Calibour on the DreamCast comes to mind (didn't that was a great game, guys??). The AI did adapted to the player when he overused a special move.

However, I think that was just a little table entry that said: "If player used move "FistFromBehind" x time in the last y attacks, avoid having a distance of z from him.

A NN could lern this, and a genetic algorithms could lern the same stuff as well, but why do the painfull job to train the network when a simple, hand-written table will do at least as well?

Besides that: In my university days, I was into all that neuronal network stuff. I used it for a race AI game once (to compress input parameters from the game engine), and I've used it with little success on OCR-jobs.

What I've learned: Neuronal networks _can_ do a great job to learn, but you have no control what they learn. I've seen NN's in training to totally ignore the input set but to learn details and propabilities of my random number generator instead :-)


To start a discussion it would be a good idea if you first get a rough idea how AI and neuronal networks work, what they can do and what not. I suggest you start with a plain two inputs, one output neuron and try to figure out what boolean operation can be trained in. (That's something you could actually try out!).

Then we can talk about the complexity of game AI, analyze the inputs of a GA or NN to a problem of your choice and find out why a hand written logic will in 99.9% of all cases be better.

Unless we talk about go or chess games, the challenge in AI is usually not to make a good AI player but to make one that is not perfect while ot feels like fake at the same time.

Nils

[akramsulaiman@hotmail.com]

thanks Nils for the speedy reply.

could you tell me what the training process for the NN consists of? and what kind of difficulties you would encounter whilst doing so? are there any risks involved?

AND FOR DAVES information;
i am not doing this report because im a gamer or am learning game designing. iam being tested on my ability to gather information. im studying to be an mech.engineer..

regards

[Nils Pipenbrinck]

It seems like you're not a programmer / student who actually needs the information for a technical project.

You're wasting my time, and you should have told me(us) what you really want: Improving your social engineering skills.

I answer to questions because I'm interested in knowledge exchange, and I don't like to be misused as a involuntarily test-person for your information gathering test.

[akramsulaiman@hotmail.com]

my project is very technical. if it wasant i wouldnt be visitng forums and looking for information. would i?? i could just copy and paste some literature from the net... but obv i want to do research and gather information from people like you who are in the industry and are uptodate. so please dont feel misused and if your still upto it answer my last few questions.
regards!!

[Nils Pipenbrinck]

No way, mate.

Why should I waste my time on you while you don't even have the guts to proof-read and correct your typing mistakes in your post.

I'm not a native english speaker, and I doubt you are. No problem for me, but I at least try to the best understandable english I can. You did so as well in your seeder post, and you failed miserable in your answer.



Hint for the future: Tell the people what your questions are all about. There's nothing bad in a information gathering test per se, but let everyone know about it. And please don't post such garbage english if you can do better.

Do you really expect me (and all the other readers) to invest my time in reading your project if you don't even have the guts to correct the obviouos typos?

What I type here is for free, usually I get paid for my suggestions. Keep that in mind for the next time.

Nils