Probabilistic Reasoning in Multiagent Systems: A Graphical Models Approach

Rating:
Summary: interesting & clearly written
Review: This book can be placed in the area of research in AI that brings single agent, centralised techniques to a distributed multi-agent context. The central idea of the multi-agent paradigm is to solve complex problems with a collection of autonomous and possibly distributed agents.

In the first 5 chapters this book gives a thorough understanding of exact inference in Bayesian networks.

In the 6th chapter Y. Xiang introduces Multiply Sectioned Bayesian networks (MSBN), a knowledge representation formalism for probabilistic inference in multi-agent systems. He clearly informs the reader of the constraints that are associated with MSBNs and how they are the unevitable consequence of a few high level choices.

Some of the choices are:
- the beliefs of the agents are represented by probabilities
- the internal representation of an individual agent is a DAG
- the least amount of communication between agents possible

Some of the contraints that follow from these choices are:
- a hypertree agent organisation which prevents agents from communicating with any other agent
- only communication between agents on variables they share between their local models

In subsequent chapters the author introduces algorithms for cooperative, distributed probabilistic inference. First how to compile an MSBN to a linked junction forest (= the multi-agent version of a junction tree) through moralization, triangulation, and the construction of linkage trees. Then, how to perform the actual probabilistic inference in such a linked junction forest.

In the 9th chapter algorithms are shown that allow to verify whether a structure does not violate the constraints imposed by the MSBN paradigm.

Finally, in the last chapter Xiang gives an overview of all the possible extensions and future work, such as dynamic formation, learning, negotiation etc.

The book is clearly written and very understandable, even for people with little knowledge of probabilistic reasoning. In my opinion this is because of the clear and not unnecessarily complicated language and because throughout the entire book the same example is used (monitoring of digital circuits ).
A few points of critique are that some more space could have been devoted to possible applications and related work.

To conclude, a very interesting and clear book on a new and promising paradigm, suited for everyone interested in Bayesian networks and multi-agent systems.