An Adaptive Generative Representation for Evolutionary Computation

James Montgomery and Daniel Ashlock
Submitted to CEC 2017

Abstract PDF eprint

This study is the second using real-coded representation for problems usually solved with a discrete coding. The adaptive generative representation is able to adapt itself on the fly to prior parts of the construction of an object as it assembles it. In the initial study the ability of the representation to take user supplied or problem supplied biases that change its behavior was demonstrated but not explored. In this study the bias is used to change the way evolution explores a fitness landscape for both an RFID antenna design problem and small instances of the traveling salesman problem. \aside{James yet to summarize his results here; will do so over the weekend}. For the traveling salesman, a simple inverse-distance bias for the adaptive generative representation causes a large improvement in performance over a random key representation in 99 of 100 instances studied.