Surrogate fitness via factorization of interaction matrix

Abstract

We propose SFIMX, a method that reduces the number of required interactions between programs and tests in genetic programming. SFIMX performs factorization of the matrix of the outcomes of interactions between the programs in a working population and the tests. Crucially, that factorization is applied to matrix that is only partially filled with interaction outcomes, i.e., sparse. The reconstructed approximate interaction matrix is then used to calculate the fitness of programs. In empirical comparison to several reference methods in categorical domains, SFIMX attains higher success rate of synthesizing correct programs within a given computational budget.