Institute for Software Technology

Similar documents
On the Multicriteria Integer Network Flow Problem

Math 261 Exercise sheet 2

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009

Fundamental loop-current method using virtual voltage sources technique for special cases

6. Stochastic processes (2)

6. Stochastic processes (2)

Linear, affine, and convex sets and hulls In the sequel, unless otherwise specified, X will denote a real vector space.

HMMT February 2016 February 20, 2016

REDUCTION MODULO p. We will prove the reduction modulo p theorem in the general form as given by exercise 4.12, p. 143, of [1].

Singular Value Decomposition: Theory and Applications

Line Drawing and Clipping Week 1, Lecture 2

2.3 Nilpotent endomorphisms

2E Pattern Recognition Solutions to Introduction to Pattern Recognition, Chapter 2: Bayesian pattern classification

FACTORIZATION IN KRULL MONOIDS WITH INFINITE CLASS GROUP

Economics 101. Lecture 4 - Equilibrium and Efficiency

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

20. Mon, Oct. 13 What we have done so far corresponds roughly to Chapters 2 & 3 of Lee. Now we turn to Chapter 4. The first idea is connectedness.

Perron Vectors of an Irreducible Nonnegative Interval Matrix

Perfect Competition and the Nash Bargaining Solution

EEL 6266 Power System Operation and Control. Chapter 3 Economic Dispatch Using Dynamic Programming

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems

Stanford University CS254: Computational Complexity Notes 7 Luca Trevisan January 29, Notes for Lecture 7

Combining Constraint Programming and Integer Programming

Expected Value and Variance

COMPLEX NUMBERS AND QUADRATIC EQUATIONS

Introductory Cardinality Theory Alan Kaylor Cline

Graph Reconstruction by Permutations

Engineering Risk Benefit Analysis

The Geometry of Logit and Probit

10.34 Numerical Methods Applied to Chemical Engineering Fall Homework #3: Systems of Nonlinear Equations and Optimization

APPENDIX A Some Linear Algebra

Lecture 17 : Stochastic Processes II

On a direct solver for linear least squares problems

Société de Calcul Mathématique SA

Every planar graph is 4-colourable a proof without computer

Feature Selection: Part 1

More metrics on cartesian products

Generalized Linear Methods

Volume 18 Figure 1. Notation 1. Notation 2. Observation 1. Remark 1. Remark 2. Remark 3. Remark 4. Remark 5. Remark 6. Theorem A [2]. Theorem B [2].

Linear Regression Analysis: Terminology and Notation

a b a In case b 0, a being divisible by b is the same as to say that

MMA and GCMMA two methods for nonlinear optimization

find (x): given element x, return the canonical element of the set containing x;

Homework Assignment 3 Due in class, Thursday October 15

Lecture 5 Decoding Binary BCH Codes

Appendix B. The Finite Difference Scheme

n α j x j = 0 j=1 has a nontrivial solution. Here A is the n k matrix whose jth column is the vector for all t j=0

Chapter 1. Probability

8.1 Arc Length. What is the length of a curve? How can we approximate it? We could do it following the pattern we ve used before

The Minimum Universal Cost Flow in an Infeasible Flow Network

Problem Set 9 Solutions

LECTURE V. 1. More on the Chinese Remainder Theorem We begin by recalling this theorem, proven in the preceeding lecture.

Statistical pattern recognition

MAT 578 Functional Analysis

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal

763622S ADVANCED QUANTUM MECHANICS Solution Set 1 Spring c n a n. c n 2 = 1.

The Expectation-Maximization Algorithm

Lecture 3: Probability Distributions

On the Operation A in Analysis Situs. by Kazimierz Kuratowski

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

Matrix Approximation via Sampling, Subspace Embedding. 1 Solving Linear Systems Using SVD

Bayesian predictive Configural Frequency Analysis

Notes on Frequency Estimation in Data Streams

Foundations of Arithmetic

Problem Do any of the following determine homomorphisms from GL n (C) to GL n (C)?

G4023 Mid-Term Exam #1 Solutions

CHAPTER 4 MAX-MIN AVERAGE COMPOSITION METHOD FOR DECISION MAKING USING INTUITIONISTIC FUZZY SETS

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg

Outline and Reading. Dynamic Programming. Dynamic Programming revealed. Computing Fibonacci. The General Dynamic Programming Technique

Learning Theory: Lecture Notes

CS 331 DESIGN AND ANALYSIS OF ALGORITHMS DYNAMIC PROGRAMMING. Dr. Daisy Tang

Lecture 4: November 17, Part 1 Single Buffer Management

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X

Gravitational Acceleration: A case of constant acceleration (approx. 2 hr.) (6/7/11)

CS286r Assign One. Answer Key

Complete subgraphs in multipartite graphs

Drago{ CVETKOVI] Mirjana ^ANGALOVI] 1. INTRODUCTION

Report on Image warping

A new construction of 3-separable matrices via an improved decoding of Macula s construction

Some Consequences. Example of Extended Euclidean Algorithm. The Fundamental Theorem of Arithmetic, II. Characterizing the GCD and LCM

Polynomial Regression Models

Be true to your work, your word, and your friend.

Module 9. Lecture 6. Duality in Assignment Problems

Transfer Functions. Convenient representation of a linear, dynamic model. A transfer function (TF) relates one input and one output: ( ) system

A 2D Bounded Linear Program (H,c) 2D Linear Programming

Lecture 7: Gluing prevarieties; products

DIFFERENTIAL FORMS BRIAN OSSERMAN

Week3, Chapter 4. Position and Displacement. Motion in Two Dimensions. Instantaneous Velocity. Average Velocity

10-701/ Machine Learning, Fall 2005 Homework 3

CHAPTER 17 Amortized Analysis

The Multiple Classical Linear Regression Model (CLRM): Specification and Assumptions. 1. Introduction

Lecture 16 Statistical Analysis in Biomaterials Research (Part II)

and problem sheet 2

A TRACTABLE APPROACH TO PROBABILISTICALLY ACCURATE MODE ESTIMATION

Affine transformations and convexity

ÉCOLE POLYTECHNIQUE FÉDÉRALE DE LAUSANNE

Physics 5153 Classical Mechanics. Principle of Virtual Work-1

Kernel Methods and SVMs Extension

Structure and Drive Paul A. Jensen Copyright July 20, 2003

On the set of natural numbers

Transcription:

- Dagnoss I - Insttute for Software Technology Inffeldgasse 16b/2 A-8010 Graz Austra 1

References Skrptum (TU Wen, Insttut für Informatonssysteme, Thomas Eter et al.) ÖH-Copyshop, Studenzentrum Stuart t Russell und Peter Norvg. Artfcal Intellgence - A Modern Approach. Prentce Hall. 2003. 2

Motvatng Example MBD a b c d M1 * M2 * M3 A1 + A2 + f g f * 3

Motvatng Example MBD M1 a 3 6 * A1 10 b 2 12 c 2 d f 3 M2 * 6 + A2 3 12 M3 * 6 + f g 4

Motvatng Example MBD M1 a 3 6 * A1 10 b 2 12 M2 4 + c 2 6 * A2 10 d 3 12 f 3 M3 * 6 + f g 5

Omn-drectonal Robot 6

Applcaton Domans 7

Prncples MBD System Descrpton (Model) Dagnoss Physcal System Desered Behavor Dscrepancy Observed Behavor 8

Requrements Prncples MBD Model (Component-Connecton-Behavor Model) Powerful Computer Benefts General Methodology Easy to mantan Easy adaptable to other problems Cost reducton 9

Defntons 1. Dagnoss System: A dagnoss system (SD,COMP) conssts of a system descrpton SD,.e., a set of FOL sentences descrbng the components behavor and the system structure, and a set of dagnoss components COMP. Alexander - Felferng and Gerald Stenbauer 1 0-10

Example: AND gates Defntons and(c) ( ab(c) out(c)=n 1 (C) n 2 (C)) and(a 1 ) and(a 2 ) out(a 1 )=n 1 (a 2 ) n 1 out n 1 & & n 2 n 2 out a 1 a 2 11

Defntons 2. Dagnoss: Let (SD,COMP) be a dagnoss system and OBS a set of observatons. A set COMP s a dagnoss ff SDUOBSU{ ab(c) CCOMP\ }U{ab(C) C } s consstent. 12

Example: AND gates Defntons OBS={n 1 (a 1 )=true n 2 (a 1 )=true n 2 (a 2 )=true out(a 2 )=false} true true n 1 n 2 & a 1 out true n 1 n 2 & out false a 2 13

Proposton 1. A dagnoss exsts for (SD,COMP,OBS) OBS) ff SDUOBS s consstent. Proof: If SDUOBS s nconsstent, then obvously t s mpossble for all COMP to fulfll the dagnoss condton. So there exsts no dagnoss. On the other hand f SDUOBS s consstent at least COMP s a dagnoss. 14

Proposton 2. {} s a dagnoss for (SD,COMP,OBS) OBS) ff SDUOBSU{ ab(c) CCOMP} s consstent. 3. Every superset of a dagnoss s a dagnoss. 4. If s a dagnoss for (SD,COMP,OBS), then for each C, SDUOBSU{ ab(c) CCOMP\ } ab(c ) 15

Proposton Proof: If ={} the result s vacuously. Suppose then that ={C 1,,C k } and that the proposton p s false. Then there exsts a C such that SDUOBSU{ ab(c) CCOMP\ } ab(c ). From the defnton of follows that there must be a logcal Model M L wth the property M L SDUOBSU{ ab(c) CCOMP\ } ab(c ). M L Now we can conclude ab(c ) whch s n M contradcton wth our ntal L assumpton C. 16

Proposton 5. s a dagnoss for (SD,COMP,OBS) ff SDUOBSU{ ab(c) CCOMP\ } { ( } s consstent. 17

Defnton 3. A conflct set for (SD,COMP,OBS) s a set COCOMP such that SDUOBSU{ ab(c) CCO} s nconsstent. A conflct set s mnmal f no proper subset s a conflct set. 18

Proposton 6. COMP s a dagnoss for (SD,COMP,OBS) ff s a mnmal set such that COMP/ s not a conflct set. 19

Defnton 4. Suppose C s a collecton of sets. A httng set for C s a set HU SCS such that H SØ for each SC. A httng set s mnmal f no proper subset s a httng set. 20

Theorem 7. COMP s a (mnmal) dagnoss for (SD,COMP,OBS) ff s a (mnmal) httng set for the collecton of conflcts set. Proof: (1) By proposton 6 COMP\ s not a conflct set for (SD,COMP,OBS). Hence, every conflct set contans an element of, so that s a httng set for the collecton of conflct sets. (2) We now show that COMP\ s no conflct. If t s a conflct set would not ht t, contradctng the fact that s a httng set. 21

Computng Httng Sets F collecton of conflcts 1.Let D represent a growng dag. Generate a node whch wll be the root of the dag. 2.Process the nodes n D n breath-frst order. To process a node n: a. Defne H(n) ) to be the set of edge labels l on the path n D from root to node n. b. If for all xf, x H(n)Ø then label n by. Otherwse, label n by where s the frst member of F whch x H(n)=Ø. c. If n s labeled by a set F, for each σ, generate a new downward arc labeled wth σ. Ths arc leads to a new node m wth H(m)=H(n)U{σ}. The new node m wll be processed after all nodes n the same generaton as n have been processed. 3.Return the resultng dag D. 22

Prunng Rules Reusng nodes: Ths algorthm wll not generate a new m as a descendant of node n. There are two cases to consder: 1. If there s a node n n D such that H(n )=H(n) U{σ}, then let the σ-arc under n pont to ths extng node n. Hence, n wll have more than one parent. 2. Otherwse, generate a new node m at the end of ths σ-arc as descrbed n the basc HS-DAG algorthm. Closng: If there s a node n n D whch s labeled by and H(n )H(n) then close the node n. A label s not computed for n nor any successor nodes are generated. 23

Prunng Rules Prunng: If the set s to label a node n and t has been used prevously, then attempt to prune D as descrbed n the followng: 1. If there s a node n whch has been labeled by the set S of F where S, then relabel n wth. For any n S \, the edge under n s no longer allowed. The node connected by ths edge and all ts descendants are removed, except those nodes wth another ancestor whch s not beng removed. Note that t ths step may elmnate the node whch h s currently processed. 2. Interchange the sets S and n the collecton. Note that ths has the same effect as elmnatng S from F. 24

Example HS-DAG F={{a,b},{b,c},{a,c},{b,d},{b}} {b {a {b {b}} n0: {},{a,b} b n2: {b},{a,c} a c n3: {a,b}, n5: {b,c}, Alexander -- Felferng and Gerald Stenbauer 25

Drawback HS-DAG Need to know or compute conflct sets n advance Idea: Compute conflct set ncrementally when they are requred by the HS-DAG algorthm Theorem Prover: TP(SD,CH,OBS) denotes a theorem prover call returnng a (not necessarly mnmal) conflct set f one exsts,.e., SDUOBSU{ ab(c) CCH} s nconsstent, and otherwse. 26

Computng Dagnoses Dagnose(SD,COMP,OBS) OBS) 1.Generate a pruned hs-dag D for the collecton F of conflct sets for (SD,COMP;OBS) as descrbed prevously, except that whenever, n the process of generatng D a node n of D needs access to F to compute ts label, label that node by TP(SD,COMP\H(n),OBS). 2.Return {H(n) n s a node of D labeled by }. 27

Example 1 Bt Full Adder X1 X2 A1 O1 A2 OBS: A=1, B=0, C n =1, S=1, C out =0 28

Example Dagnose n0: {}, TP(SD,COMP,OBS) OBS) {X1,A1,A2,O1} X1 A1 A2 O1 n1: {X1}, TP(SD,COMP/{X1},OBS) n4: {O1}, TP(SD,COMP/{O1},OBS) {X1,X2} X1 X2 n2: {A1}, TP(SD,COMP/{A1},OBS) {X1,X2} X1 n5: {A1,X1}, TP(SD,COMP/{A1,X1},OBS) X2 n3: {A2}, TP(SD,COMP/{A2},OBS) {X1,X2} n6: {A1,X2}, TP(SD,COMP/{A1,X2},OBS) {X1,A2,O1} X1 X2 n7: {A2,X1}, TP(SD,COMP/{A2,X1},OBS) n9: {O1,X1}, TP(SD,COMP/{O1,X1},OBS) OBS) n10: {O1,X2}, TP(SD,COMP/{O1,X2},OBS) n8: {A2,X2}, TP(SD,COMP/{A2,X2},OBS) OBS) 29

Conflcts 1 Bt Full Adder X1 X2 C 2 C 1 A1 O1 A2 C 3 30

Multple Dagnoss Canddates Problem: How to dstngush between several dagnoses canddates (dscrmnaton)? Idea: Use addtonal measurements? Addtonal measurements e e s are.e. costly. How to select the most valuable addtonal measurement? 31

Measurement Selecton Defnton 5: A dagnoss for (SD,COMP,OBS) OBS) predcts ff SDUOBSU{ab(C) C }U{ ab(c) CCOMP\ }.e., on the assumpton that the components of are all faulty, and the remanng components are all functonng normally, the system behavor must hold. Proposton 8: A dagnoss for (SD,COMP,OBS) predcts ff SDUOBSU{ ab(c) CCOMP\ } 32

Measurement Selecton Theorem 9: Suppose every dagnoss of (SD,COMP,OBS) predcts one of,. Then: 1. Every dagnoss whch predcts s a dagnoss for (SD,COMP,OBSU{}). 2. No dagnoss whch predcts s a dagnoss for (SD,COMP,OBSU{}). 3. Any dagnoss for (SD,COMP,OBSU{}) whch s not a dagnoss for (SD,COMP,OBS) OBS) s a strct t t superset of some dagnoss for (SD,COMP,OBS) whch predcts. Any new dagnoss resultng from the new measurement wll be a strct superset of some old dagnoss whch predcted. 33

Measurement Selecton Corollary 10: Suppose that {} s not a dagnoss for (SD,COMP,OBS). Then under the assumpton of theorem 9, any new dagnoss arsng from the new measurement wll be a multple fault dagnoss. Corollary 11: Suppose that {} s not a dagnoss for (SD,COMP,OBS). OBS) Then under the assumpton of theorem 9, the sngle fault dagnoses for (SD,COMP,OBSU{}) OBSU{ are precsely those of (SD,COMP,OBS) whch predct. 34

Example Measurement Selecton a 3 b c d M1 * A1 10 2 12 2 f 3 M2 * 5 + A2 3 12 M3 * + f g ={{M1},{A1},{M2,A2},{M2,M3}} {A1} {M2 A2} {M2 M3}} ={{M1M2A2}{M2M3A1}{M2A1A2}} {{M1,M2,A2},{M2,M3,A1},{M2,A1,A2}} 35

Next Measurement Pont Gven: dagnoss canddates (mnmal dagnoses and ther superset), fault probabltes for each component p(c), possble measurements x =v k where x denotes the quantty and v k a value. R k canddates whch reman f x s measured to be v be v k S k canddates whch x must be v k U canddates whch do not predct a value for x R k =S k UU and S k U =Ø 36

Next Measurement Pont The best measurement s one whch mnmzes the expected entropy of canddate probabltes resultng form measurement: H e ( x ) m k1 p( x v k ) H ( x v k ) where v 1,.,v m are possble values. 1 m p 37

Next Measurement Pont p ( x v ) p ( S ),0 p ( U ) m k1 p d k ( ) k p( U C k ), p( S p( C) k ) k S k p (1 C COMP\ d k ( ), p( C)) p( U ) U k p d ( ) Assume: Each v k s equal lkely ff a canddate does not predct a value x,.e., k =p(u )/m 38

Next Measurement Pont ) ( ) ( ) / ( ) ( ) ( e e k k x H H x H m U p S p v x p )) ( ln( ) ( ) ( ) ( ) ( k k n e e e v x p v x p x H ) ( ln ) ( )) ( ln( ) ( 1 k k k e U p U p n U p U p values predcted of number ln )) ( ln( ) ( S n m m U p U p )) ( ( mn )) ( ( mn values predcted number of, e e k x H x H S n 39

a b c d f 3 Example Measur. Select. M1 * M2 x 1 A1 2 10 2 3 3 * M3 * x 2 x 3 + A2 + 12 f g p(m1)=p(m2)=p(m3)=p(a1)=p(a2)=0.1 (M2) (M3) (A1) (A2) 1 40

Dagnoss p( ) x 1 x 2 x 3 M1 0.06561 4 6 6 M1,M2 0.00729 4 6 6 M1,M2,M3M2 M3 0.0008100081 - - - M1,M2,M3,A1 0.00009 - - - M1,M2,M3,A1,A2 0.00001 - - - M1,M2,M3,A2 0.00009 - - - M1,M2,A1 0.00081-6 6 M1,M2,A1,A2M2 A2 0.0000900009 - - 6 M1,M2,A2 0.00081 - - 6 M1,M3 0.00729 4 6 6 M1,M3,A1 0.00081-6 6 M1,M3,A1,A2 0.00009-6 - M1,M3,A2 0.00081 4 6 - M1,A1 0.00729-6 6 M1,A1,A2 0.00081-6 6 M1,A2 0.00729 4 6 6 41

Dagnoss p( ) x 1 x 2 x 3 A1 0.06561 6 6 6 A1,M2 0.00729 6 6 6 A1,M2,M3 0.00081 6 - - A1,M2,M3,A2 0.00009 6 - - A1,M2,A2 0.00081 6-6 A1,M3 0.0072900729 6 6 6 A1,M3,A2 0.00081 6 6 - A1,A2 0.00729 6 6 6 M2,M3 0.00729 6 4 8 M2,M3,A2 0.00081 6 4 - M2,A2 0.00729 6 4 6 42

Lne X p(x) X 1 S 1[4] 0.08829 S 1[6] 0.10539 U 1 0.0117101171 X 1 =4 0.094145 X 1 =6 0.111245 X 2 S 2[4] 0.01539 S 2[6] 0.18639 U 2 0.00361 X 2 =4 0.017195 X 2 =6 0.188195 X 3 S 3[6] 0.19368 S 3[8] 0.00729 000 U 3 0.00442 X 3 =6 0.19589 X 3 =8 0.00950 43

Example Measurement Selecton X 1 X 2 X 3 Entropy -0.458637-0.381701-0.360562 44

Problem Computng Measurements Prevous algorthm fts not for large systems, use of supersets Practcal Soluton Use only computed dagnose canddates, no subersets 45

Revsed Algorthm D set of dagnoses for (SD,COMP,OBS) p( x v k where v k ) p d D cond ( ) ( ) SD OBS v,, v undef 1 k ab( C) C COMP H ( x ) p( x vk ) ln( p( x vk )) v k \ ( x v k ) cond( ) Search for mn H(x ) 46

Example Measurment Selecton Lne X p(x) X 1 X 1 =4 0.06561 X 1 =6 0.0801908019 X 2 X 2 =4 0.01458 X 2 =6 0.13122 X 3 X 3 =6 0.13851 X 3 =8 0.06561 X 1 X 2 X 3 Entropy -0.381071-0.328138-0.452532 47

Thank You! Alexander -- Felferng and Gerald Stenbauer 48