Evolution Strategies apply mutation and recombination operators in order to create their offspring. Both operators have a different role in the evolution process: recombination should combine information of different individuals, while mutation performs a kind of random walk to introduce new values. In an ES these operators are always applied together, but their different roles suggest that it might be better to apply them independently and at different rates. In order to do so the ES has been split into two levels. The resulting Modular Evolution Strategy consists of a population of local optimizers and a distributed population manager. Both parts have their own specific role in the optimization process. As a result of its modularity this method can be adapted more easily to specific classes of numerical optimization problems, and introduction of adaptive mechanisms is relatively easy. A further interesting aspect about this algorithm is that it does not need any global communication, and therefore can be parallelized easily. Many problems can be expressed as numerical optimization problems. Especially when the dimension of the input space and the number of local optima is high these problems tend to be very difficult. In order to obtain an efficient solver one has to gather information regarding the function to be optimized. Evolution based learning can be used to obtain this information. This paper contains results obtained with the Modular Evolution Strategy and compares these results to those obtained with other evolution based method. The results look promising.

Ordinary Differential Equations (acm G.1.7), Problem Solving, Control Methods, and Search (acm I.2.8)
Problem solving (heuristics, search strategies, etc.) (msc 68T20)
CWI
Department of Computer Science [CS]

van Kemenade, C.H.M. (1995). A two-level evolution strategy : balancing global and local search. Department of Computer Science [CS]. CWI.