Re: Bag of Tricks: Choosing actions via the matching law

Falk Hueffner <[EMAIL PROTECTED]> wrote:
> It doesn not appear to be sensible to try X with the same
> probability if it worked in 1 of 3 cases, or in 100000 of 300000 cases

Yeah, at first glance, it doesn't seem like it makes sense, but that's
the way it seems to happen in the real world

Of course, there are modifiers. Animals have multiple learning
systems, each with it's own threshold for minimum number of instances.
For the gustatory (taste) learning system, you need only one trial to
learn a rock-solid association. For visual and auditory, it's like 3.
For visual and auditory for rabbits and people with Alzheimer's, it's
closer to 600

The reason why is that a)the real world is *really* sparse data and
b)in the real world, bad decisions kill you

Taste learns in one trial because you don't want to eat a poison plant
3 times to learn the plant was poisonous (as opposed to some other
reason such as time of day, where you ate it, cleanliness of fork,
etc.). In auditory and visual (not sure if this is one learning system
or two), you don't want to take too long to learn that the sound of
rustling grass behind you might mean there's a fox or eagle or
something ready to eat you

Many actions don't get a chance to get tried 2,000 times. Most you
don't want to wait too long before learning the association

If you have 500 spells and weapons to choose from, the data quickly
overwhelms you and it takes too long to scientifically try everything.
So you make a preference or decision in just a few trials. Of course,
in the real world, we're rarely told that something is a +5 weapon
doing 1-10+5 damage. That sort of info in games does help a bit :)

-b, who doesn't know nearly as much as he pretends to