Defining Implicit Objective Functions for Design Problems

Abstract

The ability of evolutionary algorithms and related search techniques to explore a varied space of solutions with efficiency and often surprising innovation makes them useful tools for design. This typically requires the explicit definition of a goal or objective function and so has been ideally suited to engineering optimisation tasks. For many design problems however, and particularly for those of great complexity, it is difficult to specify such a goal in advance. Design and creativity themselves, particularly in a social context, are often seen as processes of guided, but open exploration. Steels has shown that effective languages can be generated without an external measure of quality by allowing robots to speak and evaluate each other in an environment. Such approaches have been incorporated into genetic algorithms by allowing the objective to change over time.

The method presented here generates an objective function for a genetic algorithm (GA) based only on labelled or classified examples of previous designs, and thus requires no explicit statement of objectives. It uses supervised learning to reduce the dimensionality of initial parameters, in which only the features relevant to the original classification apply. Fitness is then determined by a distance measurement within this space.

This is demonstrated in the design of office floor plans, in which the location and orientation of many desk units form a complex spatial configuration. It is a problem of sufficient complexity that no consensus on measures of fitness currently exists. In practical use the designer need not specify one, but simply classify plans. The grid representation is generic, and can just as well apply to other domains such as urban form. Methods for improving the mapping are detailed, and the method is tested to evolve solutions as defined by various sets of archetypal plans.

Title: Defining Implicit Objective Functions for Design Problems

Author: Sean Hanna

Publication:Proceedings of GECCO '07: Genetic And Evolutionary Computation Conference | full text (PDF)

Year: 2007

D.O.I: doi.acm.org/10.1145/1276958.1277355

Tags: Machine Learning Sean Hanna architecture fitness evaluation genetic algorithm