Knowledge Based Nonuniform Crossover

We pres ent a new "knowledge-based nonuniform crossover" opera tor for genet ic algorit hms that generalizes uniform crossover. We ext end this to "Dynamic knowledge-based nonuniform crossover ," which const antly updates the knowledge ext racted t hus far from the environment 's feedb ack on previously generated chromosomes. Knowledge-based nonu niform crossover can improve on good solutions previously obtained using ot her algorit hms. The modificat ions made by knowledge-based nonuniform crossover are ort hogonal to other changes in parameters of genet ic algorithms, and can be pursued toge ther with any other proposed improvement s. Whereas most genetic search methods focus on improving move-selection pro cedures, after having chosen a fixed move-generation mechanism , knowledgebased nonuniform crossover and dynamic knowled ge-based nonuniform crossover make the move-generation process itself time-dependent. T he same parents may give rise to different offspring at different moments in the evolut ionary pro cess, based on the past experience of the species. Simulation result s show orders of magnit ude improvement of knowledge-based nonuniform crossover over two-po int and uniform crossover relati ve to thr ee NP optimization problems : graph partitioning, soft -decision decod ing of linear block codes , and the t rave ling salesperson problem. In particular, knowledge-based nonuniform crossover has been applied to variant s of the graph partit ioning problem not easily solved with other methods, and found to imp rove t he quality of the solutions. Dynamic knowledge-based nonuniform crossover opens up the possibility of applying genetic algori thms to Incremental Optimization problems, characterized by a slow change in *Corresponding author. Electronic mail address: kt shana t op . cis . syr. edu tElectronic mail address: mohanetop.cis . syr .edu tElectronic mail address: r ankactop.cis . syr .edu 258 H. Maini, K. Mehrotra , C. Mohan, and S. Ranka problem structure over time. Dynamic knowledge-based nonuniform crossover also achieves some of the goals of diploid representat ions with adapt ive dominance, with smaller computational requirements.