2004-02-22 #4

Adaptron Test Run-4 22nd Feb, 2004

3 state (A, B & C) environment like maze with rewarding novel stimulus after any response at C.

With Interesting off and no practicing it’s pretty dull.

With practicing on but still nothing interesting. Must have looping detection on to make sure it does not practice a response witch ends up with no change in stimulus.

It behaves a little better because after a C it always responds with an a as it did last time but only because it has never succeeded in getting the same result twice so it keeps practicing it.

With Looping off but Interested on it’s not whether a loop will result that stops you from practicing but the fact that the interest level in a stimulus that you are currently practicing is low. It now looks like this.

And with Interesting On it is now motivated to repeat what it has learnt. But note at 18 it didn’t store the correct LTM structure corresponding to the real world events. It also tries a b and a c after C because it is not practicing. This means it does not validate C a Q or C b W.