Adaptive Operator Selection for Optimization

Evolutionary Algorithms (EAs) are stochastic optimization algorithms which have already shown their efficiency on many application domains. This is achieved mainly due to the many parameters that can be defined by the user according to the problem at hand. However, the performance of EAs is very sensitive to the setting of these parameters, and there are no general guidelines for an efficient setting; as a consequence, EAs are rarely used by researchers from domains other than computer science. The methods proposed in this thesis contribute towards alleviating the user from the need of defining two very sensitive and problem-dependent choices: which variation operators should be used for the generation of new solutions, and at which rate each operator should be applied. The paradigm, referred to as Adaptive Operator Selection (AOS), provides the on-line autonomous control of the operator that should be applied at each instant of the search, i.e., while solving the problem. In order to do so, one needs to define a Credit Assignment scheme, which rewards the operators based on the impact of their recent applications on the current search process, and an Operator Selection mechanism, that decides which should be the next operator to be applied, based on the empirical quality estimates built by the rewards received. In this work, we have tackled the Operator Selection problem as an instance of the Exploration versus Exploitation dilemma: the best operator needs to be exploited as much as possible, while the others should also be minimally explored from time to time, as one of them might become the best in a further moment of the search. We have proposed different Operator Selection techniques to extend the Multi-Armed Bandit paradigm to the dynamic context of AOS. On the Credit Assignment side, we have proposed rewarding schemes based on extreme values and on ranks, in order to promote the use of outlier operators, while providing more robust operator assessments. The different AOS methods formed by the combinations of the proposed Operator Selection and Credit Assignment mechanisms have been validated on a very diverse set of benchmark problems. Based on empirical evidence gathered from this empirical analysis, the final recommended method, which uses the Rank-based MultiArmed Bandit Operator Selection and the Area-Under-Curve Credit Assignment schemes, has been shown to achieve state-of-the-art performance while also being very robust with respect to different problems. iii te l-0 05 78 43 1, v er si on 1 20 M ar 2 01 1