Book Review from Kybernetes 32, no. 7/8, pp. 1185-7, 2003

Probabilistic Reasoning in Multiagent Systems: A graphical models approach

by Yang Xiang

Cambridge University Press, Cambridge, 2002, xii + 294 pp., ISBN 0-521-81308-5, hardback, 45.00

This is about intelligent decision support systems to operate with uncertain data, where the decisions supported may be in such areas as business management or medical diagnosis. This is not a new field and methods depending on graphical models known as Bayesian or belief networks have been studied for two decades and are treated in a number of texts including a standard one by Judea Pearl. These methods are for processing at one location, within a single intelligent agent. The present book extends the methods to allow multiagent operation. As it is expressed in the notes on the back cover:

The framework developed in the book allows distributed representation of uncertain knowledge on a large and complex environment embedded in multiple cooperative agents and effective, exact, and distributed probabilistic inference.

The treatment is inevitably mathematical, but the presentation is designed to make it as readily accessible as possible. Each of the ten chapters except the first ("Introduction") and last ("Looking into the Future") starts with a relatively readable "guide" or "roadmap" section. These chapters are also followed by sets of examples and the book is suitable either for self-study or as the basis of a taught course.

Use of belief networks is not the only method that could be used for decision support. They are used to set up and maintain a set of beliefs about the environment to be managed and to allow decisions to be made. An alternative would be to go directly from the observations to recommended actions without an explicit model, and this is allowed by rule-based systems with their "if-then" rules. Another possibility is the use of neural nets which can be effective without any semantic labelling at all. The use of belief networks has the big advantage of transparency, such that it is possible to trace back and analyse any decision.

The treatment of multiagent systems is developed from the established methods for the single-agent case, but the book is written so as to be self-sufficient, and the first five chapters deal with the single-agent situation, but with emphasis on aspects that will be needed in the extension to multiple agents in the second half.

To introduce the methods it is necessary to choose a specific sample environment about which beliefs are to be maintained. Reference to a chemical plant or an area of medical diagnosis would not be suitable since most readers would not have the necessary specialised knowledge. The type of environment that has been chosen is digital electronic circuitry because it seems fairly safe to assume that all readers will have acquaintance with this. (I have to admit that I had forgotten the significance of the symbols and had to check elsewhere that the shape with a straight line on its input side is an AND gate and the concave one is the OR. A very small amount of this elementary knowledge is all that is needed.)

Where the possibility of faulty operation is acknowledged, a two-input gate has four binary variables associated with it, namely its two inputs, its output, and the choice between normal and abnormal. The probability of a given output can be deduced or estimated for any permutation of the other three variables. (It can be deduced as 1 or 0 for a normal component, and values can be guessed for an abnormal one, where abnormality implies erratic rather than consistently wrong operation.) Application of BayesÆ Rule then allows estimation, from observations, of the probability that the component is faulty. Probability estimates would be stored for the 16 possible permutations of the four binary variables.

In practice it would not be feasible to monitor such a simple component as a single gate in such detail but this is the basis of the method. However, the number of variables observed in a real application will certainly be such that the number of states for which beliefs must be updated is unmanageably large. The method is rendered tractable by processing smaller sets of related variables and then combining the results, using graph-theoretic methods.

The research leading to the book was stimulated by a project for medical diagnosis called the PainULim project. (Strangely, PainULim does not appear in the subject index of the book - the description is on page 139.) The aim was to develop a system for diagnosis of patients suffering from painful or impaired upper limbs due to diseases of the spinal cord, nervous system, or both. The need for a multiagent approach seems to have arisen because the diagnosing neurologists focused attention on one system at a time.

Other application areas for belief networks, mentioned in the first chapter, include advanced robotic devices such as an automatic driver of a vehicle, analysis of operation of a firm to assist business decisions, and similar analysis in the context of intelligent tutoring. Another application area is to "intelligent houses" where the operation of the services, including perhaps the reordering of supplies, is under central control. An important application area is to systems to give warning of dangerous situations in many contexts, including earthquake prediction, and an occurrence in China is mentioned where successful prediction based on a great many indicators was responsible for saving many lives.

In the final chapter possible extensions of the theory are considered, including that to dynamic, rather than fixed, environments. The extension to operate on continuous variables is also discussed, since the theory in the book assumes discrete variables. Other work on belief networks with continuous variables, but only for single agents, is referred to and it seems clear that a forthcoming project for somebody will be to extend the principles to the multiagent case.

The enhancements of the general methods, prompted by the multiagent requirement, appear to be valuable even apart from this special aspect, as suggested by the references to "effective" and "exact" in the back-cover excerpt. My strong impression is that this is a valuable and welcome comprehensive guide to the state-of-the-art in applying belief networks.

Alex M. Andrew

Back to Reviews