Reusing Learned Functionality to Address Complex Boolean Functions

Abstract

Although it is possible to identify building blocks of knowledge created by a learning classifier system in order to reuse them to solve larger scale problems, a scaling limit was still reached in certain domains. Furthermore, it was not possible to transfer functionality from one domain to another. Initial investigations have shown that it is possible and practical to reuse learned rule sets as functions in very simple problems in the same domain. The novel work here seeks to reuse learned knowledge and functionality to scale to complex problems in the same domain and to a related domain for the first time. The past work showed that the reuse of knowledge through the adoption of code fragments, GP-like sub-trees with a depth of at most two, into the XCS learning classifier system framework could provide dividends in scaling; the technique made it possible to solve until then intractable problems like the 135 bit multiplexer. The main contribution of this investigation is that a growing set of learned functions reused in the inner nodes of a code fragment tree can be beneficial. This is anticipated to lead to a reduced search space and increased performance both in terms of instances needed to solve a problem and classification accuracy. We show that through the reuse of learned functionality at the root and leaf nodes of code fragment trees, it is possible to solve complex problems such as the 18 bit hidden multiplexer problem.