Time Control or Size Control? Reducing Complexity and Improving Accuracy of Genetic Programming Models

Complexity of evolving models in genetic programming (GP) can impact both the quality of the models and the evolutionary search. While previous studies have proposed several notions of GP model complexity, the size of a GP model is by far the most researched measure of model complexity. However, previous studies have also shown that controlling the size does not automatically improve the accuracy of GP models, especially the accuracy on out of sample (test) data. Furthermore, size does not represent the functional composition of a model, which is often related to its accuracy on test data. In this study, we explore the {\em evaluation time} of GP models as a measure of their complexity; we define the evaluation time as the time taken to evaluate a model over some data. We demonstrate that the evaluation time reflects both a model’s size and its composition; also, we show how to measure the evaluation time reliably. To validate our proposal, we leverage four well-known methods to size-control but to control evaluation times instead of the tree sizes; we thus compare size-control with time-control. The results show that time-control with a nuanced notion of complexity produces more accurate models on 17 out of 20 problem scenarios. Even when the models have slightly greater times and sizes, time-control counterbalances via superior accuracy on both training and test data. The paper also argues that time-control can differentiate functional complexity even better in an identically-sized population. To facilitate this, the paper proposes Fixed Length Initialisation (FLI) that creates an identically-sized but functionally-diverse population. The results show that while FLI particularly suits time-control, it also generally improves the performance of size-control. Overall, the paper poses evaluation-time as a viable alternative to tree sizes to measure complexity in GP