No abstract

Optimization algorithms typically operate only within a fixedsized design space, solving problems with a fixed number of parameters. However, many optimization problems allow for a variable number of components, where the optimal number may not be known a priori. These problems may be solved by using a genetic algorithm that utilizes a variable-length genome. A particular challenge when using variable-length genomes is the recombination of two parent solutions to produce meaningful children. The performances of several crossover operators are investigated and compared using a sensor placement testbed problem. It is shown that the optimal number of sensors may be determined by each operator, and that performance is improved when care is taken to preserve similarities between parent solutions. Performance may also be further improved by introducing a bias when pairing parents for recombination based on their relative genome lengths.

No abstract