Culling And Teaching In Neuro-Evolution (1997)
The evolving population of neural nets contains information not only in terms of genes, but also in the collection of behaviors of the population members. Such information can be thought of as a kind of culture" of the population. Two ways of exploiting that culture are explored in this paper: (1) Culling overlarge litters: Generate a large number of offspring with different crossovers, quickly evaluate them by comparing their performance to the population, and throw away those that appear poor. (2) Teaching: Use backpropagation to train offspring toward the performance of the population. Both techniques result in faster, more effective neuro-evolution, and they can be effectively combined, as is demonstrated on the inverted pendulum problem. Additional methods of cultural exploitation are possible and will be studied in future work. These results suggest that cultural exploitation is a powerful idea that allows leveraging several aspects of the genetic algorithm.
In Proceedings of the Seventh International Conference on Genetic Algorithms (ICGA-97, East Lansing, MI), Thomas B{"a}ck (Eds.), pp. 760-767 1997. San Francisco, CA: Morgan Kaufmann.

Paul H. McQuesten Ph.D. Alumni paul [at] mcquesten net
Risto Miikkulainen Faculty risto [at] cs utexas edu