Using a neural network to determine fitness in genetic design

Many automated design approaches require an objective function to determine the quality of a given design. Often, this function is a complex relationship between many parameters that are subjective, and their relationships are difficult to quantify. This article presents a neural network based method to solve the inverse problem of determining the designer's preferences. In the forward problem, the designer would define relative preferences and the relationship between performance attributes for a given design task. The quality of a design could then be evaluated based on these preferences. The inverse problem seeks to quantify the designer's preferences and the relationships between those preferences based on evaluations of a few candidate designs. Generally, a human designer might propose candidate designs, the designer would then rank or rate the quality of the candidate designs, and then the candidate designs are used to solve the inverse problem by training a neural network fitness function. This fitness function can then be used to evaluate and create new designs that the human designer might not conceive. This article demonstrates the approach through the design of modular robots for planetary exploration.