Studies On Instance-Based Inductive Learning Applied In Genetic Algorithm Strategy Selection

Genetic algorithm is a random search algorithm based on probability-driven, and has been successfully applied to a variety of learning tasks and optimization problems. To genetic algorithm, the design and selection of strategy (which includes algorithm follow, operators, and their control parameters) has significant impact on the performance and even convergence of the algorithm. However, there is no complete standard to follow to select a good strategy. In particular, how to determine the appropriate operators and control parameters is not easy. In the traditional genetic algorithm for function optimization problems, operator combinations and the general scope of their parameters are identified on the basis of theoretical analysis and experience, and often need to spend a lot of calculations in real time, after repeated attempts to adjustment. So it has great blindness for the strategy selection. We tried to introduce a new approach, through the evaluation tests of the training functions with different strategies, to build an instance library, extract modes from function expressions, and then generate knowledge rules, which are used to guide the choice of optimization strategy of other similar functions with the same mode, from the instance-based inductive learning. According to this method, it is natural to lower the likely cost of calculation caused by repeated trials significantly, making evolutionary computation able to absorb experience from the priori similar problems, and more intelligent to choose a suitable strategy to improve its accuracy and efficiency.In this paper, we combined and used the knowledge and methods in some related areas of inductive learning, case-based reasoning, and genetic algorithms. By the way to explore the intrinsic link between the constructions of function expressions and algorithm strategies which include the algorithm flow, operators and control parameters, proposed a method to cluster functions and select the algorithm strategies based on the expression construction, providing a practical way to explore for the induction of strategy selection rules.