|
|
|
|
|
|
|
in perpetually novel environments with sparse payoff. Overt memories in such situations necessarily involve many tangled strands, including unnecessary detours and incidentals. To tease out the relevant strands at the time payoff occurs would be an overwhelming, hardly feasible task. Over repeated trials the bucket brigade carries out this task, but in an implicit fashion. |
|
|
|
|
|
|
|
|
Generating plausible replacements for rules assigned low strength under the credit assignment algorithm is an even more daunting task than credit assignment itself. In a rule-based system, the whole process of induction succeeds or fails in proportion to its efficacy in generating plausible new rules, rules that are not obviously incorrect on the basis of experience. However, plausible is not an easy concept to pin down computationally. It implies that experience biases the generation of new rules, but how? |
|
|
|
|
|
|
|
|
I propose that the concept of plausibility is closely linked to the "schema" concept set forth in the discussion of genetic algorithms. Because the rules in a classifier system are presented by strings defined over a three-letter alphabet, {1,0,#}, we can think of the strings as chromosomes defined on three alleles. Accordingly, we can interpret the set of rules used by the classifier system as a population of chromosomes. Moreover, the strength of each rule can be interpreted directly as its fitness (though it should be noted that there are interesting variants that base fitness on strength in a less simplistic way). A genetic algorithm, then, is easily applied to such a population of rules, and, indeed, classifier systems were designed with just this objective in mind. |
|
|
|
|
|
|
|
|
In this application of the genetic algorithm, schemata serve as building blocks for rules. The usefulness of any given schema can be estimated, in the usual way, from the average observed strength of the rules that are instances of the schema in the population. Though these estimates are subject to error, they do provide an experience-dependent guideline. Both the possibility of error and role of experience are consonant with the term plausibility. As always, the genetic algorithm exploits these estimates implicitly (implicit parallelism, née intrinsic parallelism in chapter 4) rather than explicitly, but this does not affect the plausibility of the new rules generated thereby. |
|
|
|
|
|
|
|
|
It helps in understanding the evolution of a classifier system to note that simple schemata (those with few defining positions) generally have more instances than more complex schemata in a population of fixed size. From a sampling point of view, this |
|
|
|
|
|