Robust evolution strategies

Abstract

Evolution Strategies(ES) are an approach to numerical optimization that shows good optimization performance. However, according to our computer simulations, ES shows different optimization performance when a different lower bound of strategy parameters is adopted. We analyze that this is caused by the premature convergence of strategy parameters, although they are traditionally treated as “self-adaptive” parameters. This paper proposes a new extended ES, called RES in order to overcome this brittle property. RES has redundant neutral strategy parameters and adopts new mutation mechanisms in order to utilize the effect of genetic drift to improve the adaptability of strategy parameters. Computer simulations of the proposed approach are conducted using several test functions.