Pareto Optimality, GA-easiness and Deception

This paper deenes a class of spaces which are easy for genetic algorithms and hard for stochastic hill{climbers. These spaces require genetic recombination for successful search and are partially deceptive. Problems where tradeoos need to be made subsume spaces with these properties. Preliminary results comparing a genetic algorithm without crossover against one with two{point crossover support these claims. Further we show how a genetic algorithm using pareto optimality for selection, outperforms both. These results provide insight into the kind of spaces where recombination is necessary suggesting further study of properties of such spaces, and what it means to be GA{easy and hill{climbing hard.