Learning for evolutionary design

Abstract

This paper describes a technique for evolving similar solutions to similar configuration design problems. Using the configuration design of combination logic circuits as a test bed, the paper shows that combining genetic algorithms with a case-based memory leads to improved performance on sets of similar design problems. In this approach, rather than starting from scratch on each design, we periodically inject a genetic algorithm's population with appropriate partial solutions to similar previously attempted problems. Experimental results on the combinational logic design of parity checkers and adders shows that this system takes less time to provide better quality solutions to new design problems as it gains experience from solving other similar design problems. The designs generated by the combined system also tend to be more similar than those generated by a randomly initialized genetic algorithm. This implies that the system can be used for quick, high quality re-design so that when components fail or deteriorate, we can quickly regain lost or deteriorating functionality.