Computer Learning

I am trying to create some virtual animals with learning ability. My initial simplified model for their virtual brains is this: (outputstate, newbrainstate ) = F[inputstate][oldbrainstate] Where each of these states can be represented by an integer. Outputstates could represent different movements for example. Brainstates can represent combinations of emotions and short term memory. The idea is that initially the array F is filled with random pairs of numbers. e.g. F[3][11]=(132,24) If the animals actions lead to it getting hurt then the arrayvalue F that got it into this position is assigned a new random pair of numbers. So by trial and error it should learn which actions to avoid. But looking closer at this model there appear to be some problems. For example firstly there are too many possibilities, there doesn't seem to be a mechanism for the animal to generalise about consequences instead of just one particular one. Also there doesn't seem to be a mechanism for dealing with bad decisions that effect the animal a few steps ahead. Does anyone know how to improve on this model perhaps with using neural nets etc.? Would the animal require a long-term memory in able to properly learn?