Toward a Mechanics of Innovation: Lessons from Genetic Algorithms

One of the attractions of recombinative genetic algorithms--search procedures based on the mechanics of natural selection and genetics--is that the processes involved seem intuitively similar to those of human innovation, but that connection is not very helpful to the design of genetic algorithms, because the processes and dynamics of innovation are themselves not well understood. In fact, the more useful line of inquiry comes from reversing the chain of reasoning. As we learn more about the design of {\it competent} GAs---GAs that solve hard problems, quickly, reliably, and accurately---it appears that the dynamics and processes of innovation are succumbing to a more mechanistic understanding.

In this talk, we examine the implications of the design of competent GAs for the understanding of recombinative innovation processes. Starting from a brief introduction of the mechanics and effect of simple GAs and moving toward the sixfold design decomposition that has been used to develop competent GAs since 1993, a working, quantitative theory of innovation is constructed. The theory is predictive in that calculation of critical "time" and "length" scales can be combined with dimensional reasoning to determine important dimensionless ratios such as the {\it decision number} and the {\t innovation number} that help us understand when and how well a innovative process will work and whether those processes will scale to larger, more difficult situations. Scalability will be demonstrated using examples from a number of competent GAs, including the fast messy GA, the gene-expression messy GA, and the linkage-learning GA. The talk closes with more speculative remarks regarding the possibilities of a computational theory of the creative.