Representation and Hidden Bias II: Eliminating Defining Length Bias in Genetic Search via Shuffle Crossover

The traditional crossover operator used in genetic search exhibits a position-dependent bias called the dcfining-length bias. We show how this bias results in hidden biases that are difficult to anticipate and compensate for. We introduce a new crossover operator, shuffle crossover, that eliminates the position dependent bias of the traditional crossover operator by shuffling the representation prior to applying crossover. We also present experimental results that show that shuffle crossover outperforms traditional crossover on a suite of five function optimization problems.