Quote:Original post by thelurch
If all you want to do is creating an interesting sample space in which to observe the behaviour of a neural network then this sounds a bit like overkill.

Afterall, it means you first have to create the environment in which you want to test it. i.e. an environment in which your choices of {ATTACK,HIDE,FLEE,WEAPON_SEARCH,MEDECIN_SEARCH,BLOCK,IDLE,SEARCH,FOLLOW,DEAD} all make sense. For instance, how clear is the difference between HIDE,FLEE and BLOCK? and in what situations will one be suitable instead of the other.

I would suggest you start with a smaller sample space first e.g.{ATTACK, FLEE, HEAL}.

If, however, you are working in a pre-existing environment, e.g. modding a game, then your outputs could be mapped to states the way xEricx described. From what I know of fuzzy logic it's practically the same thing.

As for your inputs you need to figure out what information is important in determining the best state.
For instance, distance from opponent will probably be important, availability of cover, current health status, current weapon status, weapon oppenent is carrying, opponents health status, ... basically, think of everything that would help you in making a decision of how to act in that environment, and give that information to it.

whatever i want to do with ANN, telling me it's over killing does not help me at all. using ANN in a NPC AI it's my dissertation topic.

besides, the states i have given are only some examples to illustrate my problem...

Nevertheless your last statement seems interesting...

Quote:Original post by jolyqr
could you tell me more about the fuzzy classifier?

Briefly, a Fuzzy Classifier permits the non-binary classification of a continuous input vector based on degree of membership of discrete sets. So, for example, imagine a 1-d input of height and the discrete decomposition of height into {short, average, tall, huge}. We expect that these discrete classifications actually overlap, so that a person might be placed in both the tall and average categories, but only to a degree that sums to unity. This creates a vector of degree of membership, which is then transformed using a 'defuzzification' rule to obtain a classification.

Interestingly enough, a Fuzzy Classifier is equivalent to a limited class of B-spline basis networks and is functionally equivalent to a radial basis function (RBF) network.

There's an absolute glut of information about Fuzzy classifiers, b-spline networks and rbf networks available online. If you want a good book that covers the problem you're looking at, read "Neurofuzzy adaptive modelling and control" by Martin Brown & Chris Harris.

Good luck,

Timkin

Quote:Original post by Timkin

Quote:Original post by jolyqr
could you tell me more about the fuzzy classifier?

Briefly, a Fuzzy Classifier permits the non-binary classification of a continuous input vector based on degree of membership of discrete sets. So, for example, imagine a 1-d input of height and the discrete decomposition of height into {short, average, tall, huge}. We expect that these discrete classifications actually overlap, so that a person might be placed in both the tall and average categories, but only to a degree that sums to unity. This creates a vector of degree of membership, which is then transformed using a 'defuzzification' rule to obtain a classification.

Interestingly enough, a Fuzzy Classifier is equivalent to a limited class of B-spline basis networks and is functionally equivalent to a radial basis function (RBF) network.

There's an absolute glut of information about Fuzzy classifiers, b-spline networks and rbf networks available online. If you want a good book that covers the problem you're looking at, read "Neurofuzzy adaptive modelling and control" by Martin Brown & Chris Harris.

Good luck,

Timkin

cheers !!