Introduction to Probabilistic Reasoning. Image credit: NASA. Assignment

Size: px

Start display at page:

Download "Introduction to Probabilistic Reasoning. Image credit: NASA. Assignment"

John Shaw
6 years ago
Views:

Introduction to Probabilistic Reasoning Brian C. Williams 16.410/16.413 November 17 th, 2010 11/17/10 copyright Brian Williams, 2005-10 1 Brian C. Williams, copyright 2000-09 Image credit: NASA.

1 Introduction to Probabilistic Reasoning Brian C. Williams / November 17 th, /17/10 copyright Brian Williams, Brian C. Williams, copyright Image credit: NASA. Assignment Homework: Problem Set #8: Linear Programming, due today, Wednesday, November 16 th. Problem Set #9: Probabilistic Reasoning, out today, due Wednesday, November 24 th. Readings: Today: Review of Probabilities and Probabilistic Reasoning. AIMA Chapter 13. AIMA Chapter 14, Sections Monday: HMMs, localization & mapping AIMA Chapter 15, Sections /17/10 copyright Brian Williams,

2 Notation S, Q, R, P Logical sentences Φ Background theory (a sentence). not, and ( ), or (v), implies ( ), if and only if (iff, ). Standard logical connectives where iff if and only if. M(S), entails, Models of sentence S, entails, false. A, B, C Sets., φ niverse of all elements, empty set.,, ~, - Set union, intersection, inverse and difference. Equivalent to. V: Variable or vector of variables. V i : The ith variable of vector V. V, v i : A particular assignment to V; short form for V= v, V i = v i. V t : V at time t. V i:j : A sequence of V from time i to time j. 11/17/10 copyright Brian Williams, Notation S: States or state vector. O: Observables or observation vector. X: Mode or mode vector. S0, S1 Stuck at 0/1 mode. Prefix (k) L Returns the first k elements of list L. Sort L by R Sorts list L in increasing order based on relation R. s i ith sample in sample space. P(X) The probability of X occurring. P(X Y) The probability of X, conditioned on Y occurring. A C B) A is conditionally independent of C given B. 11/17/10 copyright Brian Williams,

Outline Motivation Set Theoretic View of Propositional Logic From Propositional Logic to Probabilities Probabilistic Inference General Queries and Inference Methods Bayes Net Inference Model-based

3 Outline Motivation Set Theoretic View of Propositional Logic From Propositional Logic to Probabilities Probabilistic Inference General Queries and Inference Methods Bayes Net Inference Model-based Diagnosis (Optional) 11/17/10 copyright Brian Williams, Multiple Faults Occur three shorts, tank-line and pressure jacket burst, panel flies off. Lecture 16: Framed as CSP. Image source: NASA. APOLLO 13 How do we compare the space of alternative diagnoses? How do we explain the cause of failure? How do we prefer diagnoses that explain failure? 11/17/10 copyright Brian Williams,

4 Due to the unknown mode, there tends to be an exponential number of diagnoses. G G Good Good Candidates with NKNOWN failure modes F1 Fn Candidates with KNOWN failure modes 1. Introduce fault models. More constraining, hence more easy to rule out. Increases size of candidate space. 2. Enumerate most likely diagnoses X i based on probability. Prefix (k) ( Sort {X i } by decreasing P(X i O) ) Most of the probability mass is covered by a few diagnoses. copyright Brian Williams, /17/10 7 in 0 A X B Y C out out 0 Idea: Include known fault modes (S0 = stuck at 0, S1= stuck at 1, ) as well as nknown. Diagnoses: (42 of 64 candidates) Fully Explained Failures: [A=G, B=G, C=S0] [A=G, B=S1, C=S0] [A=S0, B=G, C=G]... Partially Explained: [A=G, B=, C=S0] [A=, B=S1, C=G] [A=S0, B=, C=G]... Faults Isolated, No Explanation: [A=G, B=G, C=] [A=G, B=, C=G] [A=, B=G, C=G] 11/17/10 copyright Brian Williams,

Estimate Dynamically with a Bayes Filter An action is taken Posterior belief after an action State Space P(S t o 1:t ): Initial belief Posterior belief after sensing Image by MIT OpenCourseWare.

5 Estimate Dynamically with a Bayes Filter An action is taken Posterior belief after an action State Space P(S t o 1:t ): Initial belief Posterior belief after sensing Image by MIT OpenCourseWare. Localizing a Robot within a Topological Map x 1 x 2 : p(x 2 x 1,a)=.9 x 3 : p(x 3 x 1,a)=.05 x 4 : p(x 4 x 1,a)=.05 Observations can be features such as corridor features, junction features, etc. Source unknown. All rights reserved. This content is excluded from our Creative Commons license. For more information, see 5

6 Outline Motivation Set Theoretic View of Propositional Logic From Propositional Logic to Probabilities Probabilistic Inference 11/17/10 copyright Brian Williams, Propositional Logic Set Theoretic Semantics Given sentence S: models of S M(S) niverse of all interpretations () 11/17/10 copyright Brian Williams,

7 Set Theoretic Semantics: S True M(True) niverse 11/17/10 copyright Brian Williams, Set Theoretic Semantics: S not Q M(not Q) M(Q) M(not Q) M(Q) 11/17/10 copyright Brian Williams,

8 Set Theoretic Semantics: S Q and R M(Q) M(Q and R) M(R) M(Q and R) M(Q) M(R) 11/17/10 copyright Brian Williams, Set Theoretic Semantics: False M(False) φ 11/17/10 copyright Brian Williams,

9 Set Theoretic Semantics: Q or R M(Q or R) M(Q) M(R) M(Q or R) M(Q) M(R) 11/17/10 copyright Brian Williams, Set Theoretic Semantics: Q implies R M(R) M(Q) Q implies R is True iff Q entails R M(Q implies R) M(Q) M(R) 11/17/10 copyright Brian Williams,

10 Set Theoretic Semantics: P implies Q M(R) M(not R) M(Q) Q implies R is True iff Q and not R is inconsistent Q implies R is True iff M(Q and not R) φ 11/17/10 copyright Brian Williams, Axioms of Sets over niverse 1. A B B A Commutativity 2. A (B C) A (B C) Associativity 3. A (B C) A B A C Distributivity 4. ~ (~A) A ~ Elimination 5. ~ (A B) (~ A) (~ B) De Morgan s 6. A (~A) φ 7. A A Identity propositional logic axioms follow from these axioms. 11/17/10 copyright Brian Williams,

11 Outline Motivation Set Theoretic View of Propositional Logic From Propositional Logic to Probabilities Degrees of Belief Discrete Random Variables Probabilistic Inference 11/17/10 copyright Brian Williams, Why Do We Need Degrees of Belief? Given theory Φ, sentence Q. is never true. could be true. must be true. M(Φ) M(Q) M(Φ) M(Q) M(Q) M(Φ) Q = One of the valves to Engine A is stuck closed. M(Φ) M(Q) Q = All of the components in the propulsion system are broken in a way never seen before. unsatisfiable satisfiable entailment 11/17/10 copyright Brian Williams,

12 P(Φ) = P(s i ) s i Φ Degrees of Belief Probability P: Events [0,1] is a measure of the likelihood that an Event is true. Events: like propositions. Φ, Q, Φ Q Φ Q Φ Q Sample Space {s i }: like interpretations. mutually exclusive, collectively exhaustive, finest grained events. P defined over s i. Like counting weighted models in logic. 11/17/10 copyright Brian Williams, Axioms of Probability P: Events Reals 1. For any event A, P(A) P() = If A B = φ, then P(A B) = P(A) + P(B). All conventional probability theory follows. 11/17/10 copyright Brian Williams,

13 Conditional Probability: P(Q Φ) Is Q true given Φ? What is P( Q given Φ )? Φ Q Φ Q P(s j Φ) = P(s j ) / P(Φ) if s j Φ 0 otherwise = P(s j, Φ) P(Φ) 11/17/10 copyright Brian Williams, Conditional Probability: P(Q Φ) Is Q true given Φ? What is P( Q given Φ )? Φ Q Φ Q 11/17/10 copyright Brian Williams,

14 Degree of Belief Given theory Φ, sentence Q. is inconsistent. could be true. must be true. M(Φ) M(Q) M(Φ) M(Q) M(Q) M(Φ) P(Q Φ)= 0 P(Q, Φ) 1 P(Φ) 11/17/10 copyright Brian Williams, Representing Sample Space sing Discrete Random Variables Discrete Random Variable X : Domain D X = {x 1, x 2, x n } mutually exclusive, collectively exhaustive. P(X): D X [0, 1] P(x i ) = 1 Joint Distribution over X 1, X n : Y {cloudy, Domain Π D Xi sunny} P(X 1, X n ): Π D Xi [0, 1] Z {raining, dry} Notation: P(x 1, x n ) P(X 1 = x 1 ^. ^ X n = x n ) 11/17/10 copyright Brian Williams,

15 Outline Motivation Set Theoretic View of Propositional Logic From Propositional Logic to Probabilities Probabilistic Inference General Queries and Inference Methods Bayes Net Inference Model-based Diagnosis (Optional) 11/17/10 copyright Brian Williams, Types of Probabilistic Queries Let X = <S; O> Belief Assessment b(s i ) = P( S i o) Most Probable Explanation (MPE) s * = arg max P(s o) for all s D S Maximum Aposteriori Hypothesis (MAP) Given A S a * = arg max P(a o) for all a D A Maximum Expected tility (ME) Given decision variables D d* = arg max x u(x)p(x d ) for all d D D 11/17/10 copyright Brian Williams,

16 Common Inference Methods and Approximations used to Answer Queries 1. Product (chain) rule. 2. Product rule + conditional independence. 3. Eliminate variables by marginalizing. 4. Exploiting distribution during elimination. 5. Conditional probability via elimination. 6. Conditional probability via Bayes Rule. Approximations: 1. niform Likelihood Assumption 2. Naïve Bayes Assumption 11/17/10 copyright Brian Williams, Outline Motivation Set Theoretic View of Propositional Logic From Propositional Logic to Probabilities Probabilistic Inference General Queries and Inference Methods Bayes Net Inference Model-based Diagnosis (Optional) 11/17/10 copyright Brian Williams,

17 Bayesian Network Burglary Earthquake TV Alarm Nap JohnCall MaryCall P(M A, N) Input: Directed acyclic graph: Nodes denote random variables. Arcs tails denote parents P of child C. Conditional probability P(C P) for each node. Output: Answers to queries given earlier. 11/17/10 copyright Brian Williams, Burglary Conditional Independence for a Bayesian Network Earthquake TV Alarm Nap JohnCall Answer MaryCall M B, E, T, J N, A P(M N, A, B, E, T, J) = P(M N, A) A variable is conditionally independent of its non-descendants, given its parents. 11/17/10 copyright Brian Williams,

18 Computing Joint Distributions P(X) for Bayesian Networks Burglary Earthquake TV Alarm Nap JohnCall MaryCall Answer Joint Distribution: Assigns probabilities to every interpretation. Queries computed from joint. How do we compute joint distribution P(J,M,T,A,N,B,E)? 11/17/10 copyright Brian Williams, Product (Chain) Rule (Method 1) P(C S) : Smoking = no light heavy P(C=none) P(C=benign) P(C=malig) P(C, S): Smoking / Cancer = 0.96 x 0.80 none benign malig no light heavy C (Cancer) {none, benig, malig} S (Smoking) {no, light, heavy} P(S) : P(S=no) 0.80 P(S=light) 0.15 P(S=heavy) /17/10 copyright Brian Williams,

19 Product Rule + Conditional Independence (Method 2) If A is conditionally independent of C given B (i.e., A C B), then: Suppose A B, C E and B C E, find P(A, B, C E): Product rule Independence 11/17/10 copyright Brian Williams, Product Rule for Bayes Nets Burglary Earthquake X i is independent of its ancestors, given its parents: TV Alarm Nap JohnCall MaryCall Product rule Independence 11/17/10 copyright Brian Williams,

20 Computing Probabilistic Queries Let X = <S; O> Belief Assessment: b(s i ) = P(S i e) Most Probable Explanation (MPE): s * = arg max P(s o) for all s D S Maximum Aposteriori Hypothesis (MAP): Given A S a * = arg max P(s o) for all a D A Solution: Some combination of 1. Start with joint distribution. 2. Eliminate some variables (marginalize). 3. Condition on observations. 4. Find assignment maximizing probability. 11/17/10 copyright Brian Williams, Eliminate Variables by Marginalizing (Method 3) Given P(A, B) find P(A) : B {raining, dry} P(A, B) A {cloudy, sunny} P(A).2.8 P(B) 11/17/10 copyright Brian Williams,

21 Exploit Distribution During Elimination (Method 4) Burglary Earthquake Bayes Net Chain Rule: X i is independent of its ancestors, given its parents. TV Alarm Nap JohnCall MaryCall Issue: Size of P(X) is exponential in X Soln: Move sums inside product. 11/17/10 copyright Brian Williams, Compute Conditional Probability via Elimination (Method 5) Example: P(A B)?.197/.2.003/.2 P(A B) Note: The denominator is constant for all a i in A.20/.8.60/.8 B {raining, dry} P(A,B) 11/17/10 copyright Brian Williams, B {cloudy, sunny}.2.8 P(B) 21

22 Computing Probabilistic Queries Let X = <S, O> Belief Assessment: S = <S i, Y> b(s i ) = P(S i o) Most Probable Explanation (MPE): s * = arg max P(s o) for all s D S Maximum Aposteriori Hypothesis (MAP): S = <A, Y> Given A S a * = arg max P(a o) for all a D A 11/17/10 copyright Brian Williams, Outline Motivation Set Theoretic View of Propositional Logic From Propositional Logic to Probabilities Probabilistic Inference General Queries and Inference Methods Bayes Net Inference Model-based Diagnosis (Optional) 11/17/10 copyright Brian Williams,

23 Estimating States of Dynamic Systems (next week) S T Given sequence of commands and observations: Infer current distribution of (most likely) states. Infer most likely trajectories. 11/17/10 copyright Brian Williams,

24 MIT OpenCourseWare / Principles of Autonomy and Decision Making Fall 2010 For information about citing these materials or our Terms of se, visit:

Directed Graphical Models

CS 2750: Machine Learning Directed Graphical Models Prof. Adriana Kovashka University of Pittsburgh March 28, 2017 Graphical Models If no assumption of independence is made, must estimate an exponential