No abstract

0000−0002−5808−7044] , Leonardo Vanneschi 2[0000−0003−4732−3328] , Sara Silva 3,4[0000−−0001−−8223−−4799] , Illya Bakurov 2[0000−−0002−−6458−−942X] and Mario Giacobini 1[0000−0002−7647−5649]Abstract. Among the various typologies of problems to which Genetic Programming (GP) has been applied since its origins, symbolic regression is one of the most popular. A common situation consists in the prediction of a target time series based on scalar features and other time series variables collected from multiple subjects. To manage this problem with GP data needs a panel representation where each observation corresponds to a collection on a subject at a precise time instant. However, representing data in this form may imply a loss of information: for instance, the algorithm may not be able to recognize observations belonging to the same subject and their recording order. To maintain the source of knowledge supplied by ordered sequences as time series, we propose a new approach to GP that keeps instances of the same observation together in a vector, introducing vectorial variables as terminals. This new representation allows aggregate functions in the primitive GP set, included with the purpose of describing the behaviour of vectorial variables. In this work, we perform a comparative analysis of vectorial GP (VE-GP) against standard GP (ST-GP). Experiments are conducted on different benchmark problems to highlight the advantages of this new approach.

No abstract

Prediction of physiological time series is frequently approached by means of machine learning (ML) algorithms. However, most ML techniques are not able to directly manage time series, thus they do not exploit all the useful information such as patterns, peaks and regularities provided by the time dimension. Besides advanced ML methods such as recurrent neural network that preserve

We exploit evolutionary computation to optimize the handcrafted Structural Similarity method (SSIM) through a datadriven approach. We estimate the best combination of luminance, contrast and structure components, as well as the sliding window size used for processing, with the objective of optimizing the similarity correlation with human-expressed mean opinion score on a standard dataset. We experimentally observe that better results can be obtained by penalizing the overall similarity only for very low levels of luminance similarity. Finally, we report a comparison of SSIM with the optimized parameters against other metrics for full reference quality assessment, showing superior performance on a different dataset.