Evolutionary Design of Reduced Precision Preprocessor for Levodopa-Induced Dyskinesia Classifier

Cartesian genetic programming, compositional coevolution, adaptive size fitness predictors, levodopa-induced dyskinesia, approximate magnitude, energy-efficient

The aim of this work is to design a hardware-efficient implementation of data preprocessing in the task of levodopa-induced dyskinesia classification. In this task, there are three approaches implemented and compared: 1) evolution of magnitude approximation using Cartesian genetic programming, 2) design of preprocessing unit using two-population coevolution (2P-CoEA) of cartesian programs and fitness predictors, which are small subsets of training set, and 3) a design using three-population coevolution (3P-CoEA) combining compositional coevolution of preprocessor and classifier with coevolution of fitness predictors. Experimental results show that all of the three investigated approaches are capable of producing energy-saving solutions, suitable for implementation in hardware unit, with a quality comparable to baseline software implementation. Design of approximate magnitude leads to correctly working solutions, however, more energy-demanding than other investigated approaches. 3P-CoEA is capable of designing both preprocessor and classifier compositionally while achieving smaller solutions than the design of approximate magnitude. Presented 2P-CoEA results in the smallest and the most energy-efficient solutions along with producing a solution with significantly better classification quality for one part of test data in comparison with the software implementation.