next up previous
Next: Accumulate Explicit Credit-Assignment Values Up: Internal Reinforcement in Neural Previous: Internal Reinforcement in Neural

Creating a Credit-Blame map

Most machine learning problems can either be described as target value prediction or classification problems. There are a number of ways in which classification problems can be reduced to target value prediction problems. In PADO this is done by evolving ``discriminator'' programs for each of the classes and then orchestrating their responses. This means that for each program in PADO, the value it should return is ordinal (i.e. if v is the right value to return, then a program that says v-1 is better than a program that says v-2). Having collapsed these two parts of machine learning into one view, let us consider an abstract input to output mapping to be learned by the neural programs.





Eric Teller
Tue Oct 29 14:55:57 EST 1996