Quote:Original post by SteadtlerThe difference between the deliberation of the FSM machine and the NN is that in the FSM most of the deliberation is performed by the designer, and for the NN it is performed by the interpolation/extrapolation of the examples provided by the designer.
And yet, as we all know, one thing the computers do well is itterations of stuff really really fast. So it is very possible to train a NN or similar structure to handle far more contingencies that is possible by hand-defining cause-effect pairs as a designer.
Thats assuming that the training will converge to an acceptable behavior. The more complex the behavior you want, the less likely learning will yield an acceptable output...
But my point here was that supervised learning perform no more *thinking* or deliberation as any other AI technique, and less than some. In both the supervised NN and the FSM, almost all deliberation is performed by the designer. It is funny to note that FSM and NN originate from the same paper, a long time ago. FSM were viewed as a form of NN :)
I suppose part of my problem here is that mentally I am thinking of more than NN as part of CI. I have my brain wrapped around the topic of my proposed GDC lecture presentation which, while not necessarily CI (and definately not NN) is far more flexible and less predictable than FSM or RBS or any of the other rigid "designer-determined" systems. Some of my comments, therefore, may reflect something other than what the original topic was about. Sorry.
It's interesting, but perhaps understandable, that everybody seems to take the "black box" problem so seriously. However, there are many CI techniques developed especially for producing (more) human-understandable data structures.
One example that comes to mind is EDDIE, which evolves decision trees (which are usually relatively easy to understand for humans) using genetic programming. It has so far only been used for financial forecasting, but could be adapted to e.g. control game agents:
http://www.bracil.net/finance/Eddie/Overview/
A related algorithm is KB-NEAT, where human-designed rules can be incorporated into evolutionary neural networks:
There are many other projects which aim to address the same problem. The main reason we are not seeing them used in the computational intelligence and games community (of academic researchers) is that most of us are not really aware of the strong need for this. Again a matter of missing communication.
Having said this, I think that the most promising use for CI techniques in games is not in tuning control routines for tasks for which it would be relatively straightforward to develop controllers manually anyway, but rather in enabling completely new types of games and game features. This would probably require designing the game with the CI algorithm in mind. One example is the automatic content creation I've been working on some prototypes of, another is the online-learning used in the proof-of-concept game NERO:
http://nerogame.org/
I'm discussing the automatic content creation idea in a recent interview:
Quote:Original post by InnocuousFox Hey Alex... how funny is it that he is referring to your site in this thread?
Not very! He's the interviewee :-)
Quote:Original post by Sneftel Do you find that flippant, passive-aggressive jabs at those you consider ignorant tend to make people more or less entrenched?
Don't blame me, I'm just following the tone set by the topic title!
Join us in Vienna for the nucl.ai Conference 2015, on July 20-22... Don't miss it!
Quote:Original post by togelius Having said this, I think that the most promising use for CI techniques in games is not in tuning control routines for tasks for which it would be relatively straightforward to develop controllers manually anyway, but rather in enabling completely new types of games and game features. This would probably require designing the game with the CI algorithm in mind. One example is the automatic content creation I've been working on some prototypes of, another is the online-learning used in the proof-of-concept game NERO
Thats sensible enough. I wish NERO and others would follow the opposite tho, i.e. establishing the feature and then demonstrating the usefulness of the method, rather than making the method become the feature.
Quote:alexjc Don't blame me, I'm just following the tone set by the topic title!
You are not following anything. You press me into establishing the discussion, and then you dont contribute anything. Dont put your lack of rhetoric on me.
