Size versus robustness in evolved sorting networks: is bigger better?

Abstract

We are interested in discovering how and why circuits developed using evolutionary methods tend to be more robust than hand designed ones. To this end, we compare evolved circuits to known, minimal sorting networks. We introduce a new, size-independent metric, called bitwise stability (BS), which measures how well a network performed when subjected to real-world types of errors. In particular, we examine stuck-on-one, stuck-on-zero and passthrough errors, a generalization of a short circuit. Networks were evolved using tree structured and linear encoded chromosomes. We found that evolution improves bitwise stability and that tree structures tend to confer more bitwise stability than linear structured chromosomes. We discuss how the size of a sorting network affects its robustness and our discovery that bigger does not necessarily mean better.