Promoting Modularity in Evolutionary Design
Hod Lipson, Jordan B. Pollack, Nam Pyo Suh · 2001
Abstract Evolutionary design systems apply principles inspired from biological evolution to automate machine design. These systems have been shown to generate simple designs for simple tasks — but their practical ability to scale up to higher complexities remains questioned. One of the keys to accomplishing higher-level evolutionary design is the ability of the process to identify and reuse knowledge discovered at lower levels, thus scaling its search capacity. One way to capture this knowledge is in the form of reusable building blocks — modules. In this paper we define modularity and discuss several approaches to promoting modularity in evolutionary design systems. In particular, we propose a new mechanism that can enhance modularization. This mechanism is based on the observation that designs that exhibit modularity have higher adaptability and consequently have better survival rates under changing requirements. Contrary to other techniques, this is a weak (indirect) formulation that docs not require representation of partial solutions or definition of a genotype from which a design is developed. We demonstrate this principle on an abstract general design problem on which modularity can be statistically quantified.