Predict the performance of ge with an ACO based machine learning algorithm

The quality of the evolved solutions of an evolutionary algorithm (EA) varies across different runs and a significant percentage of runs can produce solutions of undesirable quality. These runs are a waste of computational resources, particularly in difficult problems where practitioners have time bound limitations in repeating runs. This paper proposes a completely novel approach, that of a Run Prediction Model (RPM) in which we identify and terminate evolutionary runs that are likely to produce lowquality solutions. This is justified with an Ant Colony Optimization (ACO) based classifier that learns from the early generations of a run and decides whether to continue or not. We apply RPM to Grammatical Evolution (GE) applied to four benchmark symbolic regression problems and consider several contemporary machine learning algorithms to train the predictive models and find that ACO produces the best results and acceptable predictive accuracy for this first investigation. The ACO discovered prediction models are in the form of a list of simple rules. We further analyse that list manually to tune them in order to predict poor GE runs. We then apply the analysed model to GE runs on the regression problems and terminate the runs identified by the model likely to be poor, thus increasing the rate of production of successful runs while reducing the computational effort required. We demonstrate that, although there is a high bootstrapping cost for RPM, further investigation is warranted as the mean success rate and the total execution time enjoys a statistically significant boost on all the four benchmark problems.