Learning Heuristics by Evolutionary Algorithms with Variable Size Representation

In this paper we present an Evolutionary Algorithm (EA) that learns good heuristics from a given set of basic operations, i.e. local search search algorithms. The heuristics are given by sequences of these operations. The length of a sequencea depends on the considered problem instance, i.e. on the number of basic operations that have to be used for constructing an eecient heuristic. The sequences of dynamic length are modelled by a variable size representation in our EA. We apply the learning model to minimization of Ordered Binary Decision Diagrams (OBDDs) which are the state-of-the-art data structure in CAD for ICs. OBDDs are very sensitive to the chosen variable ordering, i.e. the size may vary from linear to exponential. The eeciency of this approach is demonstrated by experiments. By our experiments it turns out that a variable chromosome length improves the performance of our learned heuristic.