Improving the Efficiency Of Genetic Programming for Classification Tasks Using a Phased Approach
Darren M. Chitty · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
Genetic Programming (GP) uses Darwinian evolution to generate algorithms for tasks such as classification and symbolic regression. However, a drawback is the interpreter used to evaluate candidate programs adding significant computational cost. Hence, many studies have sought to improve the speed of GP primarily via parallelism. However, efficiency can also offer considerable performance gains. GP has recently been applied to combinatorial optimisation using a phased approach (Phased-GP) whereby programs are evolved piecemeal avoiding reinterpretation of sub-programs. This method was found to be highly effective and efficient compared to standard GP. This paper investigates if a similar effect is observed when using phased GP to incrementally build classifiers. Tested upon known real-world classification problems, an efficiency saving of up to 98% can be achieved with a speedup of 70 fold and no significant loss of classification accuracy. Moreover, the method can be easily used within any existing high performance parallel GP models.