The Use of Linkage Learning in Genetic Algorithms

This report investigates the use of Linkage Learning in Genetic Algorithms (GAs) to try and determine why it is worthwhile using this technique to improve the power of GAs. This report first introduces GAs and how they go about converging on optimal solutions with comparison to other search algorithms. Then it discusses what linkage learning is and several of the techniques used to perform linkage learning. The GA from ‘Learning Linkage’ [1] is then analysed. A re-implementation of the analysed GA is then tested, modified and experimented on to gain greater understanding of how the GA learns linkage and what are the motivating factors behind the improvement in linkage. From experimentation on the re-implementation it is concluded that performing crossover using recipient genomes that are the complement to the optimal building block and the truncated selection of only optimal building block donors are two of the factors that motivate improvement in linkage. Leading to the conclusion that Harik’s linkage learning GA (LLGA) may not be able to learn both linkage and fitness at the same time. Further experimentation, which tests for both improvement in linkage and fitness is suggested, to support the conjecture that Harik’s LLGA cannot learn both linkage and fitness simultaneously. Finally the conclusions from the experimentation are combined with the research on other GAs to answer the following questions: 1) How can functionally dependent genes be aligned adjacently in LLGAs? 2) What conditions are required for a LLGA to learn linkage? 3) Why is important for linkage to increase?