February 16, 2008 16:56 00139 International Journal of Neural Systems, Vol. 18, No. 1 (2008) 1–17 c  World Scientific Publishing Company AN EXPERIMENTAL STUDY OF HYBRIDIZING CULTURAL ALGORITHMS AND LOCAL SEARCH TRUNG THANH NGUYEN and XIN YAO ∗ The Centre of Excellence for Research in Computational Intelligence, and Applications (CERCIA), School of Computer Science, University of Birmingham, Birmingham, B15 2TT, United Kingdom ∗ X.Yao@cs.bham.ac.uk www.bham.ac.uk In this paper the performance of the Cultural Algorithms-Iterated Local Search (CA-ILS), a new contin- uous optimization algorithm, is empirically studied on multimodal test functions proposed in the Special Session on Real-Parameter Optimization of the 2005 Congress on Evolutionary Computation. It is com- pared with state-of-the-art methods attending the Session to find out whether the algorithm is effective in solving difficult problems. The test results show that CA-ILS may be a competitive method, at least in the tested problems. The results also reveal the classes of problems where CA-ILS can work well and/or not well. Keywords : Global optimization; continuous optimization; Cultural Algorithms; Iterated Local Search; meta-heuristic. 1. Introduction Cultural Algorithms-Iterated Local Search (CA-ILS) (see Ref. 1) is a new continuous optimization algo- rithm hybridizing local search with Cultural Algo- rithms (see Ref. 2). In Ref. 1 we have carried out some experiments to study the efficiency of CA-ILS compared with that of its predecessor: Cultural Algo- rithms. The results show that, at least in most tested functions, CA-ILS can perform equally or better than its predecessors. What we would like to know further is whether CA-ILS is a competitive method in solving new chal- lenging test problems (as described in Section 2.2 below) or not, compared to current state-of-the-art methods in the field. In addition, we also would like to understand how well CA-ILS could perform in solving problems with other special properties as: • Having noises in fitness • Non-continuous • Non-differentiable • Having narrow global optima basins • Having ill-conditioned form • Having global optima on bounds This paper attempts to answer these questions by taking some experimental study on the performance of CA-ILS in solving modern multimodal test func- tions proposed in the Special Session on Real- Parameter Optimization of the 2005 Congress on Evolutionary Computation. The rest of this paper is organized as follows: Sec- tion 2 provides some background information about the problem domain. Section 3 briefly introduces the general procedures of CA-ILS. Section 4 mentions 20 chosen test functions and 11 reference algorithms used in our study. Section 5 describes the test set- tings, and Section 6 presents the experimental results of CA-ILS in different classes of problems, com- pared to the reference algorithms and discussions. The results are further investigated in Section 7 to find out whether various properties of modern test 1