Improving search performance of linear genetic programming based image recognition program synthesis by redundancy-removed recombination

This paper propose a new recombination method, named redundancy-removed recombination, for linear genetic programming based image recognition program synthesis. The redundancy-removed recombination produces an offspring (by conventional crossover or mutation), and then adopts a canonical transformation to convert the offspring into its canonical form, in which it can be verified whether it has been evolved before (redundant). If the offspring is redundant, it is prohibited and recombination is repeated until non-redundant offspring, which has never be born in the evolutionary search, is produced. Experimental results show that the use of the redundancy-removed recombination improved the performance of evolutionary search; it converged to the global optimum faster than the use of conventional recombinations. Also we found that the redundancy-removed recombination can construct longer programs and concentrate on those areas, whereas the conventional ones cannot.