Beyond Encoding: Rotationally Invariant Operators for Evolutionary Computation

Ideally the character of an evolutionary search algorithm should be predicated on the shape of the underlying search landscape. Unfortunately, a number of operators used in evolutionary computation are very sensitive to rotation of the encoding space. The potential bias of the orientation of the encoding for these operators indicates that both the alignment of the encoding as well as the search problem influence the performance of search algorithms using these operators. This paper explores two potential mechanisms for removing rotational bias in a recombinative operator (2-point crossover) as well as a mutative operator as applied to realencoded function optimization.