Customer Reviews

Statistical Multisource-Multitarget Information Fusion

5 star:		(1)
4 star:		(1)
3 star:		(0)
2 star:		(1)
1 star:		(0)

 

 

Mahler's book is an outstanding effort to lay down a unified framework for data and information fusion. Admittedly it is a difficult book, very comprehensive, probably too applied for mathematicians and too mathematical for engineers. But it contains a very well thought out and technical vision of what it takes to process vast amounts of data, especially data of disparate types, and extract useful and correct information. Written by someone who has worked in the field for many years, I believe this book is a guide to the future of data and information fusion.

It is somewhat pretentious to write a review of this book unless one is a very established researcher in information fusion. On the other hand, such a landmark book deserves more than two reviews on amazon.com. The reception shown by the reviewer Cassandra is not uncommon, and I hope that by writing this review I can contribute towards making more researchers give the ideas of this book the consideration and scrutiny they deserve.

Mahler's aim is very ambitious: The development of a general theory that "unifies much of information fusion under a single probabilistic, in fact Bayesian, paradigm". This theory, finite set statistics (FISST), has been a major driving force behind research in target tracking and data fusion during the last decade. Therefore, to be updated on what is "hot" in these areas in the year 2011, one must be familiar with this book.

Mahler's book is ambitious because the development of such a general Bayesian theory facilitates a pursuit of Bayes-optimal estimation methods. This is of particular importance to data association problems, which the bulk of the book (chapters 9-17) is concerned with. It seems reasonably clear that truly optimal solutions to typical data association problems are computationally infeasible, but Mahler proposes very principled approximations of the optimal solutions, which may or may not be of practical utility.

There is no doubt that the theory behind these approximations is beyond most engineer's education. This book is definitely less theoretical than Mahler's previous book, but one must nevertheless be familiar with topics such as functionals and measure theory in order to understand what is going on. As an absolute minimum, any reader of this book should feel confident dealing with multivariate pdf's and nonlinear filtering (as presented in e.g. Ristic et al: "Beyond the Kalman Filter").

It is therefore hardly a surprise that many researchers simply don't understand the contributions of this book. Perhaps as a consequence of this, Mahler met som resistance when his results were undergoing peer review. As a response, he published the reviewers' comments together with his own comments in a conference paper called "Bayesian versus `plain-vanilla Bayesian' multitarget statistics". Anyone reading this book may also want to read that conference paper.

I may dare to suggest that scepticism against FISST has been prevalent in two very different camps. The first camp, which was addressed in the mentioned conference paper, consists of researchers who advocated developments similar to those of Mahler, but lacking the same theoretical depth. The other camp consists of more conservative researchers who prefer to stay with traditional methods for data association and state estimation developed in the 70s and 80s. For the practitioner, the important question is whether Mahler's new methods are more reliable and efficient than the traditional methods. It is also important to understand whether the traditional methods also can be said to approximate the theoretically optimal solutions given by FISST. In particular, one may want to have a qualified opinion on whether the approximations underlying Mahler's methods are more dramatic or less dramatic than the approximations underlying traditional tracking methods.

Mahler is more of a mathematician than an engineer, so the book contains no simulation results to confirm the utility of the proposed methods. However, Chapter 16 refers to several publications reporting successful implementations both on simulated and on real data. The reader will have to study these carefully in order to understand which merit (if any) that Mahler's methods have over traditional methods. I will encourage anyone who undertakes such a study to read these papers with an open, but critical mind. He or she should ask himself/herself questions such as: Is the SNR representative for realistic applications? Are tuning parameters such as the false-alarm rate suitable? Is comparison with the appropriate alternative methods included? Do the performance measures actually measure what matters to the reader?

Such questions may possibly be answered in these journal and conference articles, but I cannot find adequate answers in this book, and for that reason I will give it 4 stars instead of 5 stars. However, even if all of Mahler's practical methods (such as the PHD/CPHD filter etc.) were demonstrated to be useless, I believe that the underlying theory discussed in this book will still be able to establish itself as the fundamental framework of target tracking and sensor fusion for the foreseeable future. More generally, I think that researchers in many other branches of statistics would benefit from being familiar with Mahler's work. I definitely recommend this book.

I delayed buying this book because I was concerned that it would be like his earlier (co-authored) volume, which I found essentially unreadable. But I broke down and bought it. Sadly I find it even less readable than the first. Seemingly needless levels of abstraction are elaborated with little if any motivation. And the vocabulary is often arcane. Derivations based on this theory and nomenclature are hard to understand, and formulae seem to appear out of nowhere. This provides little insight to the reader. I love theory more, and applications less, than most practitioners, so I am saddened to see a presentation style so little attuned to reader needs. My regret is all the more acute since (by another approach) I know there is something very useful useful here.