The present invention relates to techniques for solving diagnosis problems, and more particularly, to a system and method using an algorithmic method for solving diagnosis problems.
Technological advances have led to the conception of a variety of complex devices and systems. When a system fails, a technician or user typically fixes the system by identifying and replacing (or fixing) the failed component. To ease this process, a variety of model-based diagnostic tools have been developed that assist a user in identifying the failed component.
A diagnosis problem arises when a system's actual behavior fails to match the expected behavior, thereby exhibiting symptoms (anomalies). System diagnosis is then the task of identifying faulty components that are responsible for anomalous behavior. To solve the diagnosis problem, one typically finds the minimal set of faulty components that explain the observed symptoms. The most disciplined technique to diagnosis is termed “model-based” because it employs knowledge of devices' operations and their connectivities in the form of models. The model-based approach reasons from first principles and affords far better diagnostic coverage of a system than traditional rule-based diagnosis methods, which are based on a collection of specific symptom-to-suspect rules.
The diagnosis process starts with identifying symptoms that represent inconsistencies (discrepancies) between the system's model (description) and the system's actual behavior. Each symptom identifies a set of conflicting components (i.e., conflict set) as initial candidates. A minimal diagnosis set is the smallest set of components that intersects all conflict sets. The underlying general approach in different model-based diagnosis approaches can be described as a two-step “divide-and-conquer” technique wherein finding the minimal diagnosis set is accomplished in two steps: 1) Generating conflict sets from symptoms; and 2) Calculating minimal diagnosis set from the conflict sets. As shown in
However, there are major drawbacks in the current model-based diagnosis techniques in efficiently performing the above two steps that severely limit their practical application to many systems of interest. First, existing conflict-generating algorithms are all based on various versions of the constraint propagation method and truth maintenance systems. The problem with these methods is that not only do they need exponential time, but they also require exponential memory to be implemented. Therefore, these methods cannot handle realistic systems with a large number of components. Second, in order to find the minimal diagnosis set, current model-based diagnosis techniques rely on algorithms with exponential computational costs and thus, are highly impractical for application to many systems of interest.
Thus, a continuing need exists for a system that allows a user to efficiently identify a minimal diagnosis set for complex systems with a large number of components.
The present invention relates to a method for calculating a minimal Hitting Set from a collection of sets. The method includes an act of using a computer system to analyze a system having a collection of sets to derive the minimal Hitting Set from the collection of sets. The computer system is configured to perform multiple acts to determine the minimal Hitting Set. For example, the computer system is configured to perform acts of receiving information regarding the collection of sets from the system; creating an incidence matrix of the collection of sets; creating nodes from the incidence matrix such that each node is a node in a search tree and has a label that includes a sub-matrix of the incidence matrix, and two disjoint subsets of the columns of the incidence matrix; calculating an upper bound and calculating a lower bound for each label, where the upper bound and lower bound define a range of solutions for the minimal Hitting Set; and determining whether each label has any child nodes or is a leaf node without any child nodes. If the label does not have any child nodes, then stopping the process and designating the label as a leaf node. If the label does have a child node, then splitting the label to create two new labels. Finally, repeating acts of calculating and determining until all labels have been designated as leaf nodes, with information gathered from leaf nodes determining a solution of the Hitting Set, thereby allowing a user to identify the Hitting Set.
In the act of calculating the minimal Hitting Set, the minimal Hitting Set is a minimal diagnosis set and the collection of sets is a collection of conflict sets. In this aspect, the method for calculating a minimal Hitting Set calculates a solution for a diagnosis problem.
In the act of creating an incidence matrix, the incidence matrix is created as a binary matrix denoted as m×n binary matrix A having columns and rows. The columns of matrix A are labeled by numbers 1,2, . . . ,n, and any subset of these columns are denoted as a subset of {1,2, . . . ,n}.
In the act of creating nodes, each node has a label (λ) in a form of λ=(M, Tin, Tout). Additionally, M is a sub-matrix of matrix A, and Tin and Tout are disjoint subsets of the columns of matrix A. Tin denotes a set of columns of matrix A considered as part of the minimal diagnosis set. Alternatively, Tout denotes a set of columns of matrix A that are considered not as part of the minimal diagnosis set.
The act of calculating an upper bound further comprises acts of calculating a value for the upper bound of λ=(M, Tin, Tout) and a calculating an upper bound set, by performing the following acts:
In the act of calculating a lower bound, a value for the lower bound of λ=(M, Tin, Tout) is equal to the size of Tin plus k/L. k denotes the number of rows of the matrix M and L denotes the maximum weight of the columns of M.
In the act of determining whether each label has any child nodes or is a leaf node without any child nodes, a label is determined to be a leaf node if any of the following are true:
In another aspect, the act of splitting the label to create two new labels further comprises an act of splitting label λ=(M, Tin, Tout) into label λ0 and label λ1. Label λ0 is defined as:
λ0=(Remove—0 [M, j], Tin, Tout∪{{tilde over (j)}}).
j denotes a column of M with a maximum weight and {tilde over (j)} is the corresponding column in the original matrix. Remove—0 [M, j] denotes a sub-matrix of M obtained by deleting the jth column of M. Label λ1 is defined as:
λ1=(Remove—1 [M, j], Tin∪{{tilde over (j)}}, Tout).
Additionally, j denotes a column of M with a maximum weight and {tilde over (j)} is the corresponding column in the original matrix. Remove—1 [M, j] denotes a sub-matrix of M obtained by deleting the jth column of M and deleting all rows of M that correspond with non-zero components of that column.
In yet another aspect, the act of splitting a label further comprises acts of:
In yet another aspect, the method further comprises an act of applying a Branch-and-Bound technique to the incidence matrix A to find an optimal solution for the problem, comprising acts of:
In another aspect, the present invention also includes a system that is configured perform the acts above.
Finally, as can be appreciated by one skilled in the art, the present invention also includes a computer program product. The computer program product includes computer-readable instruction means stored on a computer-readable medium for causing a computer to perform the operations and acts described herein.
The objects, features and advantages of the present invention will be apparent from the following detailed descriptions of the various aspects of the invention in conjunction with reference to the following drawings, where:
The present invention relates to a solution for diagnosis problems, and more particularly, to a system and method using an algorithmic solution for diagnosis problems. The following description is presented to enable one of ordinary skill in the art to make and use the invention and to incorporate it in the context of particular applications. Various modifications, as well as a variety of uses in different applications will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to a wide range of embodiments. Thus, the present invention is not intended to be limited to the embodiments presented, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
In the following detailed description, numerous specific details are set forth in order to provide a more thorough understanding of the present invention. However, it will be apparent to one skilled in the art that the present invention may be practiced without necessarily being limited to these specific details. In other instances, well-known structures and devices are shown in block diagram form, rather than in detail, in order to avoid obscuring the present invention.
The reader's attention is directed to all papers and documents which are filed concurrently with this specification and which are open to public inspection with this specification, and the contents of all such papers and documents are incorporated herein by reference. All the features disclosed in this specification, (including any accompanying claims, abstract, and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise. Thus, unless expressly stated otherwise, each feature disclosed is one example only of a generic series of equivalent or similar features.
Furthermore, any element in a claim that does not explicitly state “means for” performing a specified function, or “step for” performing a specific function, is not to be interpreted as a “means” or “step” clause as specified in 35 U.S.C. Section 112, Paragraph 6. In particular, the use of “step of” or “act of” in the claims herein is not intended to invoke the provisions of 35 U.S.C. 112, Paragraph 6.
Before describing the invention in detail, first a glossary of terms used in the description and claims is provide. Next, a description of various principal aspects of the present invention is provided. Subsequently, an introduction is provided to provide the reader with a general understanding of the present invention. Finally, the details of various aspects of the present invention are provided to give an understanding of the specific details.
(1) Glossary
Before describing the specific details of the present invention, a glossary is provided in which various terms used herein and in the claims are defined. The glossary provided is intended to provide the reader with a general understanding of the intended meaning of the terms, but is not intended to convey the entire scope of each term. Rather, the glossary is intended to supplement the rest of the specification in more accurately explaining the terms used.
Conflict Set—The term “conflict set” as used with respect to the present invention refers to a set of components of the system such that the assumption of non-faultiness of all of them is not consistent with the model of the system and observations.
Hitting Set—The term “Hitting Set” as used with respect to the present invention, refers to a minimum set that intersects all conflict sets.
Incidence Matrix—The term “incidence matrix” as used with respect to the present invention refers to a 0/1 matrix whose rows are labeled by the conflict sets {C1, C2, . . . , Cm} and columns are labeled by the components of the systems {m1, m2, . . . , mn}, and the entry aij=1 if the jth component mj belongs to the ith conflict set Ci, otherwise aij=0.
Instruction Means—The term “instruction means” as used with respect to this invention generally indicates a set of operations to be performed on a computer, and may represent pieces of a whole program or individual, separable, software modules. Non-limiting examples of “instruction means” include computer program code (source or object code) and “hard-coded” electronics (i.e., computer operations coded into a computer chip). The “instruction means” may be stored in the memory of a computer or on a computer-readable medium such as a floppy disk, a CD-ROM, and a flash drive.
(2) Principal Aspects
The present invention has three “principal” aspects. The first is a system for solving diagnosis problems. The system is typically in the form of a computer system operating software or in the form of a “hard-coded” instruction set. This system may be incorporated into a wide variety of devices that provide different functionalities. The second principal aspect is a method, typically in the form of software, operated using a data processing system (computer). The third principal aspect is a computer program product. The computer program product generally represents computer-readable instructions stored on a computer-readable medium such as an optical storage device, e.g., a compact disc (CD) or digital versatile disc (DVD), or a magnetic storage device such as a floppy disk or magnetic tape. Other, non-limiting examples of computer-readable media include hard disks, read-only memory (ROM), and flash-type memories. These aspects will be described in more detail below.
A block diagram depicting the components of a diagnosis system of the present invention is provided in
An illustrative diagram of a computer program product embodying the present invention is depicted in
(3) Introduction
The diagnosis problem arises when a system's actual behavior fails to match the expected behavior, thereby exhibiting symptoms (anomalies). System diagnosis is then the task of identifying faulty components that are responsible for anomalous behavior. To solve the diagnosis problem, one must find the minimal set of faulty components that explain the observed symptoms (i.e., minimal diagnosis set).
The present invention relates to a system for solving diagnosis problems by determining the lower and upper bounds on the size of the solution. The system uses these bounds to generate a solution window for the problem. The present invention also includes a Branch-and-Bound technique that is faster than existing techniques in terms of number of operations (by exploiting the structure of the problem). Using the concept of a solution window, the Branch-and-Bound technique allows a massive reduction (pruning) in the number of branches. Furthermore, as the Branch-and-Bound proceeds, the solution window is dynamically updated and narrowed to enable further pruning.
The present invention includes a powerful, yet simple, representation for calculating the minimal diagnosis set. This representation enables mapping onto a 0/1 Integer Programming problem. Mapping onto the 0/1 Integer Programming problem enables the use of a variety of algorithms that can efficiently solve the problem for up to several thousand components. The algorithms of the present invention are significantly improved over the existing ones, enabling efficient diagnosis of large, complex, systems. In addition, this mapping enables a fast determination of the lower and upper bounds on the solution (i.e., the minimum number of faulty components).
For further clarification, specific aspects of the present invention are described below. Additionally, tests were performed to further demonstrate the efficacy of the present invention. For further illustrative purposes, the results of performance tests of the present invention are compared with the traditional and standard algorithms. As shown below, the test results demonstrate that the present invention has a performance with more than ten times improvement over the traditional approach.
(4) Details of Specific Aspects
For clarity, various details of specific aspects of the present invention will be described in separate sub-sections below. The first section describes an algorithmic approach to the diagnosis problem. The second section describes the bounds on the diagnosis. The third section provides an introduction to the Branch-and-Bound method. The fourth section describes a Branch-and-Bound method according to the present invention. The fifth section describes performance results of the method of the present invention contrasted with a traditional Branch-and-Bound method. The sixth section explains mapping other problems into the Integer Programming Problem. And finally, the seventh section includes a brief summary of the present invention.
(4.1) Algorithmic Approach to the Diagnosis Problem
To overcome the complexity of calculating a minimal diagnosis set, the present invention relates the calculation and solution of the Hitting Set problem to the solution of Integer Programming problems.
The Hitting Set Problem, also known as the Transversal Problem, is one of the key problems in the combinatorics of finite sets and the theory of diagnosis. The problem is described as follows. A collection S={S1, . . . , Sm} of nonempty subsets of a set M is given. A Hitting Set of S is a subset H of M that intersects every set in the collection S. Of course, there are always trivial Hitting Sets. For example, the background set M is always a Hitting Set. However, interest remains in minimal Hitting Sets with minimal cardinality. A Hitting Set H is minimal if no proper subset of H is a Hitting Set.
minimize x1+x2+ . . . +xn,
subject to Ax≧b, xj=0, 1, (1)
where bT=(1, . . . ,1) is a vector whose components are all equal to one. This mapping provides for the utilization of existing Integer Programming algorithms, such as commercially available Integer Programming tools.
(4.2) Bounds on the Diagnosis
The mapping of the present invention offers two additional advantages that can be exploited to develop yet more efficient algorithms. First, the mapping represents a special case of the Integer Programming Problem due to the structure of matrix A (i.e., a binary matrix) and vector b. Second, by using this mapping, the minimum number of faulty components can be determined without solving the problem explicitly. For this purpose, the 1-norm and 2-norm of vectors are defined as:
Using the above equations and vector b from the optimization problem (1), results in ∥b∥1=m and ∥b∥2=√{square root over (m)}. Because the elements of both vectors Ax and b are positive, the following two inequalities can be derived:
where ∥ denotes the norm, x denotes multiplication, and/denotes division. Since x is a binary vector, then both norms in equation (2) give the bound on the size of the solution (i.e., the number of nonzero elements of vector x). The number of nonzero elements of vector x corresponds to the minimal diagnosis set. It should be noted that depending on the structure of the problem (i.e., the 1- and 2-norm of the matrix A and m), a sharper bound can be derived from either of the norms in equation (2). The present invention derives the bounds on the solution of the problem without explicitly solving the problem. Such a priori knowledge on the size of solution can be used for developing much more efficient algorithms for the problem.
Furthermore, an upper bound for the solution size can be found using monotonicity of the Integer Programming of optimization problem (1). Monotonicity means that if x is a solution of Ax≧b and y≧x, then y is also a solution of the same system. Note that finding a 0/1 solution x for the system Ax ≧b is equivalent to finding a subset of the columns of the matrix A, such that their sum is a vector with components all equal to or greater than 1. Any such solution provides an upper bound for the optimization problem (1), since that problem solves a minimal set of such columns.
Therefore, to find an upper bound, a column C1 of A with the largest weight must first be chosen. Then, a submatrix of A is constructed by deleting the column C1 and all rows of A that correspond to non-zero components of C1. The same process is then applied to the new matrix (i.e., submatrix of A), until an empty matrix is derived. The columns C1, C2, . . . , Ct that are obtained determine a solution for Ax≧b and the number t is an upper bound for the solution of the Integer Programming problem of optimization problem (1).
There are two simple rules that assist the above algorithm in the following two extreme cases. The algorithm will likely end up with sub-matrices that these rules can be applied to. The rules are as follows:
1. If the matrix A has an all-one column, then the upper bound is equal to 1; and
2. If some row of the matrix A has weight 1, then remove that row and the corresponding column to obtain the matrix A1 and Upper_Bound [A]=1+Upper_Bound [A1], where Upper_Bound [A1] and Upper_Bound [A] denote the values of the upper bound for the corresponding matrices.
The upper bound can also be improved by a step-by-step method and in an iterative fashion. For example, the cost of the kth step in the iteration is of the order of nk (i.e., O(nk)), where n is the number of the columns of the matrix. In this example, the first few steps are practically efficient. More specifically, for a fixed k, instead of choosing the maximum weight column for the vector a1, the sum of k columns of A can be chosen, with all possible such vectors thereafter being tried.
As another application of the a priori lower bound, the cases where a high number of faulty components requires another course of action instead of usual identification of faulty components, can be separated. These cases can be separated before starting to solve the hard problem of finding the minimal Hitting Sets. Also a good lower and upper bound can determine whether the enhanced brute-force algorithm can provide a solution efficiently. As it was stated before, this algorithm has a complexity of O(nt), where t is the number of faulty components and n is the number of the columns of the matrix.
(4.3) Introduction to the Branch-and-Bound Method
The Branch-and-Bound method is one of the most common methods for solving intractable problems. In the case of the Integer Programming (IP) problem, this method traditionally begins by solving the Linear Programming (LP) relaxation of the IP (i.e., by removing the condition that the variables xj, in the optimization problem (1) of paragraph 00066, are integers).
If the optimal solution of the LP relaxation consists only of integer values, then the optimal solution of the LP relaxation will be the optimal solution for the IP problem.
Otherwise, if the IP problem is defined by a system such as the optimization problem (1), the optimal solution of the LP relaxation provides a lower bound for the IP problem. In such a case, one of the non-integer values of the optimal solution of the LP relaxation can be chosen (e.g., xj=a), with two new sub-problems being defined by adding the conditions xj≦[a] and xj≧[a]+1 to the system, where [a] denotes the integer part of a. Then, the two new sub-problems need to be solved. By continuing this procedure, the sub-problems of the original IP problem are defined. Once an integer optimal solution for the LP relaxation of a sub-problem is found, that solution gives us an upper bound for the IP problem. After finding such an integer solution, any sub-problem is eliminated whose (LP relaxation) lower bound is larger than the upper bound provided by some other sub-problem. This procedure is continued until all sub-problems are eliminated or an integer optimal solution is found for the corresponding LP relaxation problem. At the end, the optimal solution for the IP problem is the best of the optimal integer solutions of the sub-problems.
The above procedure can be generalized as follows. To begin the Branch-and-Bound method, the following acts need to be performed:
Note that in the case of LP relaxation, whenever an integer solution for a sub-problem is found, it is guaranteed that it is an optimal solution for that sub-problem. Starting with the original problem P, defined by a system like the optimization problem (1), the Mth step of the procedure provides partitions Pm,1, Pm,2, . . . , Pm,t of mutually disjoint sub-problems. For each sub-problem Pm,j, a lower bound Lm,j is found and, if possible, an integer solution and corresponding upper bound Um,j is found. Then, a decision is made as to which sub-problems Pm,j should be eliminated at this step. There are two criteria for this decision:
1. The lower bound Lm,j is larger than upper bound Um,k of some other sub-problem Pm,k; and
2. An optimal solution for the sub-problem Pm,j is found.
In the case of (2) (i.e., optimal solution is found), the record of the best optimal solution of the sub-problems is kept. Then, the act of partitioning a problem P into a collection of mutually disjoint sub-problems is applied to the remaining sub-problems and to find the lower and (if possible) upper bounds for the new sub-problems. This procedure is continued until no sub-problem remains. Then, the best of the optimal solutions of sub-problems is the optimal solution for the original problem.
(4.4) Branch-and-Bound Method of the Present Invention
The Branch-and-Bound algorithm of the present invention is based on the above methods for computing lower and upper bounds for diagnosis. The monotonicity property of the Integer Programming problem is also exploited. The rational behind this approach and its possible advantages over the standard Branch-and-Bound method, are as follows:
Before describing the method according to the present invention, a set of useful functions and notation are introduced.
Matrix A is an m×n binary matrix. The columns of A are labeled by the numbers 1, 2, . . . , n, and any subset of these columns is denoted as a subset of {1, 2, . . . , n}. Similar to the traditional Branch-and-Bound method, the algorithm of the present invention is also based on a search of the nodes of a tree. Each node of the search tree has a label of the form:
(M, Tin, Tout),
where M is a submatrix of the original matrix A, and Tin and Tout are disjoint subsets of the columns of the original matrix A. The meaning of this partition is that Tin is the set of the columns (of the original matrix A) considered as part of the solution, while Tout is the set of the columns (of the original matrix A) that are considered not to be part of the solution. Note that a solution of the optimization problem (1) can be considered as a subset of the columns of matrix A, whose addition is a vector with all non-zero components.
Following is a list of auxiliary functions and sub-routines that are used in implementing the method of the present invention. As can be appreciated by one skilled in the art, the titles and names are for illustrative purposes only and can be changed provided the operation provides the same or a similar result.
(4.4.1) Function Place Finding
Consider a node with the label (M, Tin, Tout). Since M is a submatrix of the original matrix A, every column of M corresponds with a unique column of A. As a non-limiting example, the column 1 of M corresponds with the column 3 of A and the column 2 of M corresponds with the column 7 of A, and so on. Therefore, using a placing finding function, the label of a column of M in A can be referred to without any ambiguity. In fact, given the sets Tin and Tout, for every column j of M, it is possible to find the corresponding column in the original matrix A. For illustrative purposes, this relation is denoted by the function Place_Finding. Additionally, to keep the notation simple, instead of writing Place_Finding[Tin, Tout, j], {tilde over (j)} alone may be used if the sets Tin and Tout are understood from the context.
(4.4.2) Function Remove 0 and Remove 1
As described in further detail below, a label can be split to create two new labels. For example, label λ=(M, Tin, Tout) can be split into label λ0 and label λ1. Label λ0 is defined as:
λ0=(Remove—0 [M, j], Tin, Tout∪{{tilde over (j)}}),
where j denotes a column of M with a maximum weight, ∪ denotes the union of two sets, and {tilde over (j)} is the corresponding column in the original matrix. As applied to this section, the function Remove—0 [M, j] denotes a sub-matrix of M obtained by deleting the jth column of M.
Additionally, label λ1 is defined as:
λ1=(Remove—1 [M, j], Tin∪{{tilde over (j)}}, Tout),
where the function Remove—1 [M, j] denotes a sub-matrix of M obtained by deleting the jth column of M and deleting all rows of M that correspond with non-zero components of that column.
As a non-limiting example, consider the following matrix:
Then, to compute Remove—1 [M, 2], the 2nd column of M is deleted. Since the 2nd and 4th elements (e.g., components) of this column are 1, the 2nd and 4th rows are deleted as well. The result is as follows:
For the function Remove—0 [M, 2], only the 2nd column of M is deleted. The result is as follows:
(4.4.3) Function Rule*
This function is based on some rules that simplify the process of finding the optimal solution for the optimization problem (1):
Following is a description of how the above Rules affect the labels of the nodes in the search tree, with the label of a node being denoted as (M, {Tin, Tout}).
First, the action of Rule 1 is described. If the matrix M does not have an all-one column (i.e., a column having elements all equal to “1”), then no action is performed and the label is unchanged. Otherwise, assume that the jth column is all-one. In such a case, Rule 1 changes the label to the following:
(Ø, Tin∪{{tilde over (j)}}, Tout),
where Ø denotes an empty matrix and ∪ denotes the union of two sets (i.e., the union of Tin and {tilde over (j)} (the corresponding column in the original matrix)).
In a more formal language, function Rule—1 is defined on a set of labels as follows:
if M has no all—one columns,
if the jth column is an all—one column.
To define the action of Rule 2, first a useful notation is introduced. ej denotes the unit binary vector (of weight one) with its only 1 component at jth position. The action of Rule 2 can now be described by the following function:
if M has a row
equal to ej,
otherwise.
The action of Rule 3 is described by the following function:
if M has all-one rows,
otherwise,
where M′ is obtained from M by deleting all all-one rows.
The action of Rule 4 is described by the following function:
if the jth
column of M is all-zero,
otherwise.
Finally, the following function describes Rule 5:
if M has two equal rows,
otherwise,
where the matrix M′ is obtained from M by deleting one of the equal rows.
Note that once one of these rules is applied on a label λ1=(M, Tin, Tout), and the result is the label λ2, then it may be possible to apply one of these rules to λ2, and so on. For this reason, the Function Rule* is defined on the set of labels as repeated applications of Rules—1-5, until none of the rules can be applied anymore. It is easy to show that Rule* is well-defined (i.e., the result of Rule*(λ) does not depend on the order that the functions Rule—1 and Rules—2 are applied). As a non-limiting example, consider the following label:
λ1=(M1, Ø, Ø),
where
The 2nd row of M is equal to e2 and its 3rd row is equal to e4. Therefore, when applying Rule—2 there are two possible choices. First, the 2nd row is removed, the result is the label λ2=(M2, {2}, Ø), where
The 2nd row of M2 is equal to e3, thus Rule—2 can be applied. The result is the label λ3=(M3, {2, 4}, Ø), where
Note that the 3rd column of M3 is the 4th column of the original matrix M1. Finally, Rule—1 is applied and the result is the label λ4=(Ø, {2, 3, 4}, {1, 5}) (again note that the 2nd column of the matrix M3 corresponds with the 3rd column of the original matrix M1). Therefore, in this non-limiting example,
Rule*[λ1]=(Ø, {2, 3, 4}, {1, 5}).
(4.4.4) Function Split
As described briefly above, a label can be split into two new labels. The label is split when it is determined to not be a leaf node in the search tree (i.e., Function Test Leaf as described in further detail below). A leaf node is a node that is at an end of a branch in the search tree and has no child nodes. When a node is split, it creates at least two children nodes (i.e., child nodes). The (partial) function Split is defined on the set of the labels, where the value Split [λ] is a pair (λ0, λ1) of labels. Suppose that λ=(M, T1, T2). If T1∩T2={1, . . . n}, i.e., if T1∩T2 is the set of all columns of the matrix A, then the function Split is not defined. ∩ denotes the intersection of two sets (i.e., the set of elements common between two sets). Otherwise, let j∉T1∩T2 be a column of the original matrix A which is corresponded with a maximum-weight column of the submatrix M (if there are several such columns then the first one is chosen). ∉ denotes a membership relation, where the left-hand side is not a member of the set in the right-hand side. Then, two new labels are defined by adding the jth column to the sets T1 and T2, respectively. More specifically, the two new labels are defined as follows:
λ0=(Remove—0 [M, j], T1, T2∪{{tilde over (j)}}),
λ1=(Remove—1 [M, j], T1∪{{tilde over (j)}}, T2).
Using the above formulation, the Split function is defined as:
Split[λ]=(Rule*[λ0],Rule*[λ1]).
(4.4.5) Function Upper Bound
When solving the optimization problem (1), it is helpful to identify the range of solutions for the minimum diagnosis set. The range of solutions is defined by the upper bound and the lower bound. The Upper_Bound function is defined to find the number Upper_Bound [A] as an upper bound on the solution of the optimization problem (1). This function is extended to the set of labels as follows. For the label λ=(M, T1, T2), where M is a submatrix of the original matrix A, Upper_Bound [λ] provides an upper bound for the optimization problem defined by (1), augmented by the following conditions:
xj=1, xj∈T1, and
xj=0, xj∈T2.
Applying the above, it can be shown that Upper_Bound [λ]=|T1|+Upper_Bound [M].
In the special case that M=Ø, Upper-Bound [λ]=|T1|. Note that the function Upper_Bound is applied to both matrices and labels.
(4.4.6) Function Upper Bound Set
For the label λ=(M, T1, T2), the function Upper_Bound_Set [λ] returns the set which realizes the bound Upper_Bound [λ] (i.e., the union of T1 and the set of columns of matrix M which provide the bound Upper_Bound [M]).
(4.4.7) Function Lower Bound
Like the previous function, the lower bound defined by (2) is extended to the set of labels. More specifically, for the label λ=(M, T1, T2), where M is a k×j submatrix of the original matrix A, Lower_Bound [λ]=|T1|+k/∥M∥1. In other words, a value for the lower bound of λ=(M, Tin, Tout) is equal to the size of Tin plus k/L, where k is the number of rows of the matrix M and L is the maximum weight of the columns of M. That is, each column of M has a weight and the maximum weight is the largest among these numbers.
(4.4.8) Function Test Solution
This function is defined on the set of the labels and its value is either True or False. The value of Test_Solution [(M, T1, T2)] is True if the columns in the set T1 form a solution for the optimization problem (1). Otherwise, the value of the function is False.
(4.4.9) Function Test Leaf
This function is defined on the set of the labels and its value is either True or False. Function Test_Leaf is used to determine whether or not a node in the search tree is a leaf (i.e., whether that node has any children or not). The arguments of this function include a label λ=(M, T1, T2) and a value U for the upper bound on the solution of the problem. The function is as follows:
(4.4.8) The Branch-and-Bound Algorithm of the Present Invention
The present invention uses the above functions and terminology to create a Branch-and-Bound algorithm. The algorithm is described below for matrix A: /* solves the Hitting Set problem defined by the problem (1)*/
{Rule* [(A, Ø, Ø)]}
∞ /* upper bound */
Ø
Labels − {λ}
T1
Upper_Bound [λ]
T1
Upper_Bound [λ]
Upper_Bound_Set[λ]
Split [λ]
Labels ∪ {λ0, λ1}
Upper_Bound_Set[λ]
For clarity, the following key can be used to define various functions and terminology used in the above algorithm:
As can be appreciated by one skilled in the art, the above terminology associated with each function is used for illustrative purposes only and can be changed provided the function provides the same or a similar result. Additionally, the above algorithm is a non-limiting example of pseud-code for achieving a solution and can be altered to provide the same result.
(4.5) Performance Results of the Present Invention Contrasted with a Traditional Branch-and-Bound Method
For further illustration, consider the following 10×12 exemplary matrix:
The nodes of the search tree and their labels are as follows.
(4.5.1) Node 0
For Node 0, the label of the root is λ0=(A0, Ø, Ø). In this case, apply Rule*[λ0]=λ0. The initial lower and upper bounds are Lower_Bound=2, and U=Upper_Bound=5, where the upper bound is realized with the set {3, 4, 5, 6, 8}. Moreover, Test_Leaf [λ0, U]=False, thus this node has two children (i.e., Node 1 and Node 2). To find the two children, notice that column 8 has a maximum weight, then;
Note that none of the Rules 1-5 can be applied on the matrix Remove—0 [A0, 8]. Additionally, only Rule 4 can be applied to Remove—1 [A0, 8] because the 1st column of this matrix is an all-zero vector. Applying this rule results in the following matrix:
Therefore, the labels of the children of Node 0 are
λ1=(A1, Ø, {8}), and
λ2=(A2, {8}, {1}).
(4.5.2) Node 1
The label for Node 1 is λ1. The updated lower and upper bounds are Lower_Bound=2 and U=Upper_Bound=4. The upper bound is realized with the set {3, 5, 6, 9}, and Test_Leaf [λ1, U]=False. Thus, Node 1 has two children (i.e., Node 3 and Node 4). Column 8 of the matrix A1, which is the 9th column of the original matrix A0, has the maximum weight. Then,
None of the Rules 1-5 can be applied on the matrix A3=Remove—0 [A1, 8]. Additionally, only Rule 4 can be applied to Remove—1 [A1, 8 ] because 4th column of this matrix, which is also the 4th column of the original matrix A0, is an all-zero vector. Applying this rule provides the following matrix:
Therefore, the labels of the children of Node 1 are
λ3=(A3, Ø, {8, 9}), and
λ4=(A4, {9}, {4, 8}).
(4.5.3) Node 2
The label for Node 2 is λ2. The updated lower and upper bounds are Lower_Bound2=4 and U=Upper_Bound=4. Since Lower_Bound=U and the function Test_Leaf [λ2, U]=True, this node is a leaf of the search tree. Thus, no further splitting is done with this node.
(4.5.4) Node 3
The label for Node 3 is λ3. The updated lower and upper bounds are Lower_Bound=2 and U=Upper_Bound=3. The upper bound is realized with the set {6, 11, 12}, and the function Test_Leaf [λ3, U]=False. Thus, this node has two children (i.e., Node 5 and Node 6). Column 9 of the matrix A3, which is the 11th column of the original matrix A0, has the maximum weight. Then,
None of the Rules 1-5 can be applied on the matrix Remove—0 [A3, 9] and Remove—1 [A3, 9]. Therefore, the labels of the children of Node 3 are
λ5=(A5, Ø, {8, 9, 11}), and
λ6=(A6, {11}, {8, 9}).
(4.5.5) Node 4
The label for Node 4 is λ4. The updated lower and upper bounds are Lower_Bound=3 and U=Upper_Bound=3. Since Lower_Bound=U and Test_Leaf [λ4, U]=True, this node is a leaf of the search tree.
(4.5.6) Node 5
The label for Node 5 is λ5. The updated lower and upper bounds are Lower_Bound=3 and U=Upper_Bound=3. Since the Lower_Bound=U and the function Test_Leaf [λ5, U]=True, this node is a leaf of the search tree.
(4.5.7) Node 6
The label for Node 6 is λ6. The updated lower and upper bounds are Lower_Bound=3 and U=Upper_Bound=3. Since the Lower_Bound=U and the function Test_Leaf [λ6, U]=True, this node is a leaf of the search tree.
Therefore, the minimal Hitting Set in this example is {6, 11, 12} which is produced by Node 3.
(4.6) Mapping other Problems into the Integer Programming Problem
The optimization problem formulated in system (1) not only covers the Hitting Set Problem, but it also covers several other important combinatorial optimization problems. For example, it can also be applied to the Set Covering Problem and the Vertex Cover Problem.
(4.6.1) The Set Covering Problem
The Set Covering Problem is formulated as follows. A collection S={S1, S2, . . . , Sn} of subsets of a set X={x1, x2, . . . , xm} is given. Using the Set Covering Problem, a user can find the minimum number of sets from the collection S such that their union is the background set X. In this case, the rows of the incidence matrix A are labeled by the elements x1, x2, . . . , xm and the columns by the sets S1, S2, . . . , Sn. Thus, one skilled in the art can appreciate that the solution of the optimization problem (1) is indeed the solution of the Set Covering Problem.
(4.6.2) The Vertex Covering Problem
The Vertex Covering Problem is formulated as follows. Given a graph G=(V, E), with the vertices V={v1, v2, . . . , vn} and edges E={e1, e2, . . . , em}, find a minimum size set of vertices U ⊂ V such that for every edge ej∈E, at least one of the endpoints belongs to the set U.
In this problem, the rows of the incidence matrix A are labeled by the edges e1, e2, . . . , em and the columns by the vertices v1, v2, . . . , vn. In the row which is labeled by the edge ej, there are exactly two entries equal to 1 and those correspond with the endpoints of the edge ej. As can be appreciated by one skilled in the art, the solution of the optimization problem (1) is again the solution of the Vertex Cover Problem.
For further illustration,
(4.7) Summary and Conclusions
The present invention provides a new approach to overcome one of the major limitations of the current model-based diagnosis techniques (i.e., the exponential complexity of calculation of minimal diagnosis set). To overcome this challenging limitation, the present invention describes an algorithmic approach for calculating the minimal diagnosis set. Starting with the relationship between the calculation of minimal diagnosis set and the Hitting Set problem, the present invention includes a method for solving the Hitting Set Problem, and consequently the diagnosis problem. This method is based on a powerful and yet simple representation of the problem that enables its mapping onto another well-known problem, that is, the 0/1 Integer Programming problem.
Mapping onto the 0/1 Integer Programming problem enables the use of variety of algorithms that can efficiently solve the problem for up to several thousand components. Therefore, the algorithm of the present invention provides a significant improvement over existing ones, enabling the efficient diagnosis of large, complex systems. In addition, this mapping enables a priori and fast determination of the lower and upper bounds on the solution (i.e., the minimum number of faulty components) before solving the problem. This is exploited to develop a more powerful algorithm for the problem. The new algorithm is an improvement of the well-known Branch-and-Bound method. The results of the performance of the new algorithm are presented on a set of test cases. The results show the advantage of the new algorithm over the traditional Branch-and-Bound algorithm. More specifically, the algorithm of the present invention has achieved more than 10 times the speed of standard algorithms.
This is a non-provisional application claiming the benefit of priority of U.S. Provisional Application No. 60/652,459, filed on Feb. 11, 2005, entitled, “New High Performance Algorithmic Solution for Diagnosis Problems.”
The invention described herein was made in the performance of work under a NASA contract, and is subject to the provisions of Public Law 96-517(35 USC 202) in which the Contractor has elected to retain title.
Number | Name | Date | Kind |
---|---|---|---|
5566092 | Wang et al. | Oct 1996 | A |
5808919 | Preist et al. | Sep 1998 | A |
5922079 | Booth et al. | Jul 1999 | A |
Number | Date | Country | |
---|---|---|---|
20060195302 A1 | Aug 2006 | US |
Number | Date | Country | |
---|---|---|---|
60652459 | Feb 2005 | US |