Koevoluční algoritmy a klasifikace

The aim of this work is to automatically design a program that is able to detect dyskinetic movement features in the measured patient's movement data. The program will be developed using Cartesian genetic programming equipped with coevolution of fitness predictors. This type of coevolution allows to speed up a design performed by Cartesian genetic programming by evaluating a quality of candidate solutions using only a part of training data. Evolved classifier achieves a performance (in terms of AUC) that is comparable with the existing solution while achieving threefold acceleration of the learning process compared to the variant without the fitness predictors, in average. Experiments with crossover methods for fitness predictors haven't shown a significant difference between investigated methods. However, interesting results were obtained while investigating integer data types that are more suitable for implementation in hardware. Using an unsigned eight-bit data type (uint8_t) we've achieved not only comparable classification performance (for significant dyskinesia AUC = 0.93 the same as for the existing solutions), with improved AUC for walking patient's data (AUC = 0.80, while existing solutions AUC = 0.73), but also nine times speedup of the design process compared to the approach without fitness predictors employing the float data type, in average.

machine learning, classification, evolutionary algorithm, genetic algorithm, genetic programming, cartesian genetic programming, coevolutionary algorithm, adaptive fitness predictor, dyskinesia