Scribd
Upload
10 views

ReviewOfReasoningWithProbabilisticAndDeterministicGraphicalModelExactAlgorithms

This is a review of the book: Reasoning with Probabilistic and Deterministic Graphical Models by Rina Dechter

Uploaded by

ksp_eda
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
10 views

ReviewOfReasoningWithProbabilisticAndDeterministicGraphicalModelExactAlgorithms

This is a review of the book: Reasoning with Probabilistic and Deterministic Graphical Models by Rina Dechter

Uploaded by

ksp_eda
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

Reasoning with probabilistic

and deterministic graphical


models: Exact Algorithms
(2nd ed.)

Dechter R., Brachman R., Rossi


F. Morgan&Claypool Publishers, 2019.
Type: Book,

Many problems related to learning and reasoning can make use of graphical models
where structure of the knowledge is compactly encoded into a graph. When
dependencies (or independencies) among the variables of concern are effectively
captured in the graphical form, it becomes possible to develop efficient methods for
answering queries over that information.

Though representation of a given problem into its graphical form requires expertise
related to that particular domain, methods of reasoning from graph based
structures are universal and applicable across problem domains. With this insight in
mind, the author is presenting exact algorithms for propositional graphical models
over discrete variables in this short book.

Reasoning algorithms can be based on inference or search. Bucket elimination is an


example of inference based scheme described in this book and involves the
generation of an equivalent representation of the input problem through a sequence
of deductive steps. Search based algorithms repeatedly perform a conditioning step
that fixes value of a variable and are described through AND/OR search spaces.
Towards the end, the book also covered methods that combine search and
inference.

Chapters in the book are organized with content building on earlier material, such
that a reader that follows from beginning to end in order derives maximum benefit.
Adequate introduction is given to make the book self contained, while offering an
elaborate bibliography for reference. Part of the current work is based on an earlier
book by the same author on constraint processing.

Inference is introduced with the adaptive consistency algorithm for constraint


networks. It works by processing and eliminating variables one by one, while
deducing the effect of the eliminated variable. The procedure involves a join
operation over all the relations defined on the current variable. Constraints are
partitioned into buckets based on a given variable ordering, and after all buckets
are processed without discovering any inconsistency, the method can generate a
solution in a backtrack-free manner.

Complexity of adaptive consistency is linear in the number of buckets but the time
to process each bucket depends on generating the join of all relations which is
exponential in the number of variables mentioned in a bucket. Thus efficiency of
this approach depends on finding a variable ordering that yields the smallest
induced-width of the graph, and the book suggests some greedy heuristics for this
task.

Bucket elimination can be applied to probabilistic networks, by replacing the join


and project operators of adaptive consistency with product and summation. Here
the task is referred to as belief updating and involves determining posterior
probability of singleton variables based on new evidence. An algorithm called BE-bel
is considered in much depth. To seek belief for a variable, it needs to initiate the
ordering. Complexity of BE-bel is exponential in the number of variables of the
largest bucket.

Cluster tree elimination algorithms take inference further and avoid the need to call
BE-bel multiple times to answer belief query on multiple variables. Considering
bucket elimination as a message passing algorithm along a rooted bucket tree,
these algorithms augment it with a second set of messages from bottom to top.

Search based reasoning algorithms can be executed in a linear space and are the
only choice for models with large tree width, large domains and implicit
representation. The book considers AND/OR search spaces in much detail as their
models can be exponentially smaller when compared to linear search.

A general class of spanning trees, called pseudo-trees are considered to guide the
AND/OR search. Such a tree has the property that any arc of the graph that is not
part of the tree, connect a node to one of its ancestors in the pseudo tree. Size of
the AND/OR search tree depends on the height of its pseudo tree. To generate
compact AND/OR graphs from AND/OR trees, the book presents a merge operation
that makes use of the notion of equivalence based on a universal graphical model.

Towards the end, methods to combine inference and search are discussed. Such a
scheme can be applied to graphs that contain cycles. Consistent values for variables
of the cycle-cutset are found through search, where as remaining problem can be
solved through inference.

Presentation is pleasant and explanations are lucid. All the needed basics are
described and adequate examples are provided.

You might also like