I have an idea for making a relationsship between factors and then use some of the factors as the base for the ai behaviour. Its hard to explain, but think of its as two layers; internal and input. The "input" is set by the world/game. The "internal" can not be set, only read. Example: Confidence = AmmoLeft + 2 * Health + 3 * DistanceToAlly Here, everything except Confidence is input and set by the gameengine. Confidence is then calculated based on the current state of the game. Another example: Anger = 2 * Confidence + WeaponPower Here we see another factor being defined by one input and one internal. Im not sure if there is a name for this system (maybe some variant of a neural net?) but I would like to have some pointers where I can read about it, or similiar systems.

It's called weighted sums. A method of expressing relative utility in Decision Theory. The idea is simple. Games have been doing it since the 80's. The creation of the coefficients themselves is somewhat of a mysterious art, however.

(Shameless plug... this will be one of the topics in my book "Behavioral Mathematics for Game AI" - hopefully on the shelves by GDC.)

Quote: Original post by Mizipzor
I have an idea for making a relationsship between factors and then use some of the factors as the base for the ai behaviour.

Its hard to explain, but think of its as two layers; internal and input. The "input" is set by the world/game. The "internal" can not be set, only read.

Example:
Confidence = AmmoLeft + 2 * Health + 3 * DistanceToAlly

Here, everything except Confidence is input and set by the gameengine. Confidence is then calculated based on the current state of the game.

Another example:
Anger = 2 * Confidence + WeaponPower

Here we see another factor being defined by one input and one internal.

Im not sure if there is a name for this system (maybe some variant of a neural net?) but I would like to have some pointers where I can read about it, or similiar systems.

In the past (in my own project) Ive called them 'Synthetic Attributes' because their value at any point in time is based on the values of other information.
Ive even implimented schemes where they cache the value after being run so that the same calculation doesnt have to be rerun over and over (giving the same value) -- of course then it gets more complicated requiring a mechanism to decide WHEN its time to reevaluate it again (like when the input data changes or is expected to have changed).

The caching system may be VERy useful if you have attributes like 'PathToCurrentTarget' and the evaluation function is an A* path serach against the current map situation.....


It also get more fun when the input variable/attributes themselves are Synthetic as well.

And ontop of that, for some AI systems (like the one I have been working on) you have certainty/uncertainty information about the returned value data, And some 'values' themselves are vectors of values -- like a behavioral 'cost' type (used in returns from evaluating a task which is a candidate in a solution set) which is actually a combination of risk/time/material/certainty ).