quote: Original post by Magmai Kai Holmlor
Has anyone ever run across an AI template library?

About the closest I''ve seen (and I haven''t really looked very hard) is OpenAI.

Timkin

Not sure if this is what you mean, but if you're into Genetic Algorithms check out GALib.

-Kirk

http://lancet.mit.edu/ga/

[edited by - KirkD on June 16, 2002 9:25:21 AM]

I think there''s a lot of potential for generic templated AI. How many times must a particular algorithm be implemented before someone should make a library out of it?

Eventually, implementing path finding or neural networks should be a simple matter of calling a class member function, and most programmers won''t know the implementation details. Like for Singletons. Nowadays, you don''t really have to know how Loki or Boost ensure that there is only one instance of a Singleton to use it. It has been successfully abstracted. The same thing should be done with AI.

Cédric

Yep, I realize generic vs. genetic...

GALib is generic with respect to GAs. It is highly templated and very easy to use. If GAs are an option (perhaps to evolve a worthy opponent following a rule based AI system), I'd give GALib a try.

-Kirk

[edited by - KirkD on June 17, 2002 5:42:10 PM]

quote: Original post by cedricl
Eventually, implementing path finding or neural networks should be a simple matter of calling a class member function, and most programmers won''t know the implementation details.

Certainly you could implement a template class for such algorithms, however the problem boils down to should you! Would a template library neural network be as efficient as one tailored to the environment? If you have to do a lot of tailoring, how much use is the template library in the first place? These are the sorts of questions that need to be answered before someone makes an industry standard library.

Here''s a simple scenario to consider: computing a path cost for search.

Do you:

a) provide a domain specific callback function and register that with the library; or,

b) provide a domain model (and an interface format) to the library within which it computes its cost; or,

c) something else?

Clearly a) is easier to implement than b), however b) is likely to be more efficient as the search algorithm can work directly with the domain model, rather than having to make enquiries of an external function.

I''m not, however, suggesting that we shouldn''t look toward an AITL. Taking away the grunt work from designing game AI provides more time to look at higher level, more interesting designs for game AI.

Cheers,

Timkin

quote: Original post by Timkin
Here's a simple scenario to consider: computing a path cost for search.

Do you:

a) provide a domain specific callback function and register that with the library; or,

b) provide a domain model (and an interface format) to the library within which it computes its cost; or,

c) something else?

I think that's a great argument in favor of an AI template library. Policy-based design can provide a lot of flexibility, at little or no run-time costs. In all honesty, I've read and practiced a lot more with templates than I have with AI, so I may not have sufficient experience to say if it's really worth or not. However, as you have mentionned, a library takes away from some of the grunt work, and one of the great advantages of policy-based design is the ability to write your own policies, so that if optimization is needed, it can be done.

Besides, a lot of games have really poor AI that could definitely benefit from any library that is even remotely efficient. Some games don't even feature computer-controlled bots in multiplayer matches!

EDIT Note:
Templated callback function are as efficient as regular functions:

template< int (*func)() >void f(){   int x = (*func)();}f< mycallback >();    

With a good compiler, mycallback() could even be inlined in f()!

God, I love templates

Cédric


[edited by - cedricl on June 17, 2002 9:36:49 PM]