Two learning methods for a tree-search combinatorial optimizer

Several combinatorial problems of logic synthesis and other CAD problems have been solved in a uniform way using a general-purpose tree-searching program MULT-II. Two learning methods that have been implemented to improve the program's efficiency are presented. A weighted heuristic function, used to evaluate operators, is applied during a solution tree search. The optimal vector of coefficients for this function is learned in a simplified perceptron scheme. By using the second learning method, the similarity of shapes among the solution cost improvement curves is used to define the termination moment of the search process. The amplification effect of the concurrent action of both these methods has been observed.<<ETX>>