Hitoshi Iba, Topon Kumar Paul, Yoshohiko Hasegawa: Applied genetic programming and machine learning

Applied Genetic Programming and Machine Learning provides readers with tools and insights to apply genetic programming (GP) to classical machine learning problems. After two introductory chapters, Chap. 3 presents genetic programming for regression/forecasting problems. Chapter 4 applies genetic programming to classification problems. Finally, Chap. 5 shows how GP can be enhanced by probabilistically learning from the data generated during an optimization run. The most important aspect of this book is the accompanying code/software. Examples in the book are derived from these tools. This enables the reader to actively learn the material by practicing on the examples. The second chapter is an excellent primer for application of genetic programming in a variety of machine learning problems. The authors give extensive examples for each of the machine learning problems and provide software and data files on their website. This could be excellent teaching material for a graduate level machine learning course which would like to introduce GP as one of the possible methods. In the third chapter, the authors detail an algorithm that they have developed over a number of years called STROGANOFF (sTructured Representation On Genetic Algorithms for NOnlinear Function Fitting). The algorithm combines genetic programming with a multiple regression analysis method called Group Method of Data Handling (GMDH). Each node in the tree is a transfer function (which in this particular algorithm is a quadratic polynomial of the two input variables). GP identifies the tree structure, for the regression problem where each node performs the quadratic operation on its inputs and the GMDH algorithm finds the parameters for the transfer functions. This method lies somewhere in between parametric (where the structure of the model is assumed and only parameters are fine tuned) and free-scale non-parametric methods (in which both structure and the parameters are evolved). The method assumes a structure for interaction of variables, which in