It is hard to imagine a case where the saying “don’t judge a book by its cover” would be more appropriate than for Stalph’s Analysis and design of machine learning techniques. The book is actually the author’s PhD dissertation, apparently without much modification from its original form. Most unsuspecting readers will be disappointed with the editing of the book, as minor language issues permeate the text and the arguments are presented with the usual academic brevity. More importantly, to most computing scientists, “analysis and design” is synonymous with a systematic way of designing and evaluating computational solutions, often by grouping problems into classes of difficulty. However, this book does not give us such a systematic treatment of machine learning solutions. Instead, it makes the case for using a specific class of machine learning techniques for robotic limb control. This is done by surveying machine learning techniques, developing the math to translate kinematics into learnable functions, and applying local learning methods to the problem. Experimental results, obtained through the simulation of two different scenarios, are encouraging.

The book is organized into three parts. Part 1, “Background,” includes chapters 2 through 4 and provides a well-organized overview of machine learning techniques, taking the reader from the simplest to more sophisticated models. Chapter 2 starts by laying out the fundamental goal of machine learning: approximating an unknown function by looking at training data. Using a single and compelling example, the book takes the reader through the various models making the differences between the techniques rather palpable. Chapter 3 outlines local methods, and chapter 4 details the one advocated in the book: XCSF [1]. The idea is to use a number of kernels (each with a local model) to cover the domain of the function sampled during learning, using a genetic algorithm to explore the space of possible kernel configurations.

Chapters 5 and 6, which make up Part 2, provide arguments supporting the choice of XCSF. A brief discussion of the relevant issues pertaining to the applicability of the method is given in chapter 5, while empirical evidence to illustrate and quantify the learning ability of the approach is given in chapter 6. As this is a PhD thesis, this part, and particularly chapter 5, will probably seem rather short for a general audience. The empirical evidence provided, although fairly convincing, comes from only three experiments, which again might seem insufficient for those readers wanting a more definitive argument.

Part 3 is dedicated to robotics and how the machine learning techniques described in the previous chapters can be used for robotic limb control. Chapter 7 starts with a discussion of the physics behind robot arm movement, focusing on kinematics. The solution described in the book applies to velocity, although an argument is given that the same techniques should also allow for acceleration to be taken into account. The control loop discussed in the book is somewhat limited, ignoring obstacles and other complications that may happen in a real application. Since the goal is to determine if the learning techniques can be used at all, these simplifications seem justified.

Chapter 8 is the heart of the book. It puts together all of the elements. It starts by showing how velocity kinematics can be learned (Section 8.1) by applying the learning frameworks to approximate the parameterized linear function that governs the robotic arm movement. A good discussion about how to obtain training data and some of the potential pitfalls in how to process it with a genetic algorithm are given as well. The book argues for the suitability of such an approach and why the mathematical underpinnings of the learning method apply in this case. The approach is validated through simulation of an anthropomorphic robotic arm, as was done by previous researchers.

The experiments compare XCSF against two other learning approaches on a number of simulated tasks. The results indicate that the proposed framework, with the XCSF approach, is successful. Not only can the kinematics behind the robotic arm be learned, but the error (measured as an angle) is relatively small. Moreover, XCSF seems to offer a good trade-off compared to the other learning approaches: while it takes a bit longer to learn, it yields better results and requires less memory.

Chapter 9 provides further support for the framework, by applying it to the iCub simulator, which is a much more complex robot in which arm control takes into account visual information acquired by cameras on the robot. Section 9.2 discusses how learning can be applied to control the arm and the head of the robot in this complex scenario. The experimental validation shows that, while the results drop compared to the simpler setting of chapter 8, the framework is robust enough to yield good results.