bayesian decision theory

34
Institute of Systems and Robotics ISR – Coimbra Mobile Robotics Lab Bayesian Approaches 1 jret t

Upload: belinda-lopez

Post on 01-Jan-2016

96 views

Category:

Documents


6 download

DESCRIPTION

Bayesian Decision Theory. Pattern Classification. Bayesian Decision Theory. Retrospective. Bayesian Multimodal Perception by J. F. Fereira. Bayes' theorem - Bayes rule. Knowledge of past behavior and state form prediction of current state. Non-Gaussian likelihood functions. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 11

jrettjrett

Page 2: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 22

jrettjrett

Page 3: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 33

jrettjrett

Retrospective

Bayesian Multimodal Perception by J. F. Fereira

Bayes' theorem - Bayes rule

Knowledge of past behavior and stateform prediction of current state

Non-Gaussian likelihood functions

Multimodal Sensing in human perception

Page 4: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 44

jrettjrett

Retrospective

distribution of object position unknown => flat

Noise in each modality is independent

bimodal posterior distribution = product of the unimodal distributions

Simplification: Probability distributions are Gaussian

Page 5: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 55

jrettjrett

1. Introduction to Pattern Recognition

Example:

“Sorting incoming Fish on a conveyor according to species using optical sensing”

Page 6: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 66

jrettjrett

1. Introduction to Pattern Recognition

Selecting length feature

Example: Fish Classifier

Selecting lightness feature

Page 7: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 77

jrettjrett

1. Introduction to Pattern Recognition

Example: Fish Classifier

Selecting two features and defining a simple straight line as decision boundary

Best performance but complicated classifier – will not perform well with novel patterns

Search for the optimal tradeoff between performance on the training set and simplicity

Page 8: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 88

jrettjrett

1. Introduction to Pattern Recognition

post-processing

classification

feature extraction

segmentation

sensing

input

decision

Invariant FeaturesTranslation

RotationScale

OcclusionProjective Distortion

RateDeformation

Feature Selection

NoiseMissing Features

Error RateRisk

ContextMultiple Classifiers

Pattern Recognition System

Page 9: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 99

jrettjrett

1. Introduction to Pattern Recognition

collect data

choose features

choose model

train classifier

evaluate classifier

end

start

Prior KnowledgeOverfitting

Design Cycle

Page 10: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 1010

jrettjrett

2. Continouos Features

State of nature Finite set of c states of nature (‘categories’) {1, … , c}

Prior P(j)If the state of nature is finite:

Decision rule (for c =2):Decide 1 if P(1) > P(2); otherwise decide 2

Page 11: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 1111

jrettjrett

2. Continouos Features

Feature vector x : x d the feature spacex is (for d=1) a continuous random variable x

Class(State)-conditional probability density function: p(x| j) expresses the distribution of x depending on the state of nature

Page 12: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 1212

jrettjrett

2. Continouos Features

Bayes formula(Posterior)

Evidence

Bayes Decision rule (for c =2):Decide 1 if P(1 | x) > P(2 | x) ; otherwise decide 2

Bayes Decision rule (expressed in terms of Priors):Decide 1 if p(x|1)P(1) > p(x|2)P(2) ; otherwise decide 2

Page 13: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 1313

jrettjrett

2. Continouos Features

Conditional Risk

We can minimize our expected loss by selecting the action that minimizes the conditional risk.

This Bayes decision procedure provides the optimal performance

Two-Category Classification

Bayes Risk

Decide 1 if (21-11)P(1 | x) > (12-22)P(2 | x) ; otherwise decide 2

Page 14: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 1414

jrettjrett

Discrete Features

Probabilities rather than probability densities.

Bayes decision ruleTo minimize the overall risk, select

the action I for which R(i|x) is minimum

Feature vector x can assume m discrete values

Po

sterio

rE

vide

nceR

isk

Page 15: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 1515

jrettjrett

Discrete FeaturesExample: Independent Binary Features

2 category problemFeature vector x = {x1, …, xd}T where xi = {0;1}

Page 16: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 1616

jrettjrett

Bayesian Belief Networks

Represents knowledge about a distribution.

Knowledge: Statistical Dependencies – Causal Relations among the

component variables

Knowledge from e.g. structural information

Graphical representation: Bayesian Belief Nets

node variableP(a)link

parents(of C)

children(of E)

Page 17: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 1717

jrettjrett

Bayesian Belief NetworksApplying Bayes rule to determine the probability of any configuration of variables in the joint distribution.

P(a1) P(a2)

0.739 0.261

P(c1|ak) P(c2|ak)

a1 0.3 0.7

a2 0.6 0.4 =1

=1

=1

Discrete Case:Discrete number of possible values A (e.g. 2: a={a1, a2} andcontinues-valued probabilities

Conditional Probability Table

Page 18: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 1818

jrettjrett

Bayesian Belief NetworksDetermining the probabilities of the variablesP(a) P(b|a) P(c|b) P(d|c)

A B C D

independance

Summing the full joint distribution P(a,b,c,d) over all variables other than d

E.g.: Probability distribution over d1, d2, … at D

simple split

P(b)P(c)P(d)

simple interpretation

Probability of a particular value of D

Page 19: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 1919

jrettjrett

Bayesian Belief NetworksGive the values of some variables (evidence e)

… and search to determine some particular configuration of other variables x

Page 20: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 2020

jrettjrett

Bayesian Belief NetworksExample: Belief Network for Fish

As usual: Compute P(x1 salmon) and P(x2 sea bass) Decide for the minimum expected classification error

Ex.2 Classify the fish:Known: Fish is light (c1) and caught in the south Atlantic (b2).Unknown: Time of year (a), thickness (d)

In this case D does not affect our results

Page 21: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 2121

jrettjrett

Bayesian Belief NetworksExample: Belief Network for Fish

Page 22: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 2222

jrettjrett

Bayesian Belief NetworksExample: Belief Network for Fish

And if the dependency relation is unknown?

naïve Bayes – idiot Bayes

Features are conditionally independant

After normalization:

Page 23: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 2323

jrettjrett

Compound Bayesian Decision TheoryConsecutive ’s not statistically

independent=> exploit dependence

=> improved performance

Wait for n states to emerge and make all n decisions jointly

= compound decision problem

States of nature = ((1), … , (n))T

taking one of c values {1, … , c}

Prior P()for n states of nature

Feature matrix X : =(x1, …, xn)xi obtained when state of nature was i

n observations

Page 24: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 2424

jrettjrett

Compound Bayesian Decision TheoryDefine loss matrix for the

compound decision problem. Seek decision rule the minimizes the

compound risk (optimal procedure)

Assumption: Correct = no lossErrors = equally costly

=> simply calculate P(|X) for all and select for which P(.) is

maximum.

practice: calculate P(|X) is time expensive

assumption: xi depends only on (i) not on other x or

Conditional probability density function: p(X|) for X given the true set of

Po

sterio

r join

t d

ensity

Page 25: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 2525

jrettjrett

Obrigado!

Page 26: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 2626

jrettjrett

Annex

Page 27: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 2727

jrettjrett

Book: Pattern Cl.

PrefaceCh 1: IntroductionCh 2: Bayesian Decision TheoryCh 3: Maximum Likelihood and Bayesian EstimationCh 4: Nonparametric TechniquesCh 5: Linear Discriminant FunctionsCh 6: Multilayer Neural NetworksCh 7: Stochastic MethodsCh 8: Nonmetric MethodsCh 9: Algorithm-Independent Machine LearningCh 10: Unsupervised Learning and ClusteringApp A: Mathematical Foundations

Page 28: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 2828

jrettjrett

2 Bug Algorithms 17

3 Configuration Space 39

4 Potential Functions 77

5 Roadmaps 107

6 Cell Decompositions 161

7 Sampling-Based Algorithms 197

8 Kalman Filtering 269

9 Bayesian Methods 30110 Robot Dynamics 349

11 Trajectory Planning 373

12 Nonholonomic and Underactuated Systems 401

Book: Principles of ...

Page 29: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 2929

jrettjrett

Book: Artificial ...

PrefacePart I Artificial Intelligence Part II Problem Solving Part III Knowledge and Reasoning Part IV Planning Part V Uncertain Knowledge and Reasoning Part VI Learning Part VII Communicating, Perceiving, and Acting Part VIII Conclusions

Page 30: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 3030

jrettjrett

Book: Bayesian ...

Preface xixPart I: Fundamentals of Bayesian Inference 11 Background 32 Single-parameter models 333 Introduction to multiparameter models 734 Large-sample inference and frequency properties of Bayesian inference 101Part II: Fundamentals of Bayesian Data Analysis 1155 Hierarchical models 1176 Model checking and improvement 1577 Modeling accounting for data collection 1978 Connections and challenges 2479 General advice 259Part III: Advanced Computation 27310 Overview of computation 27511 Posterior simulation 28312 Approximations based on posterior modes 31113 Special topics in computation 335Part IV: Regression Models 35114 Introduction to regression models 35315 Hierarchical linear models 38916 Generalized linear models 41517 Models for robust inference 44318 Mixture models 46319 Multivariate models 48120 Nonlinear models 49721 Models for missing data 51722 Decision analysis 541Appendixes 571

Page 31: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 3131

jrettjrett

Book: Classification ...

Preface. Foreword. 1. Introduction. 2. Detection and Classification. 3. Parameter Estimation. 4. State Estimation. 5. Supervised Learning. 6. Feature Extraction and Selection. 7. Unsupervised Learning. 8. State Estimation in Practice. 9. Worked Out Examples. Appendix

Page 32: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 3232

jrettjrett

Images

Page 33: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 3333

jrettjrett

2. Simple ExampleDesigning a simple classifier for gesture recognition.The observer tries to predict which gesture might be performed next.The sequence of gestures appears to be random.

Ten types of gestures:1. Big circle2. Small circle3. Vertical Line4. Horizontal Line5. Pointing North-West6. Pointing West7. Talk louder8. Talk more quiet9. Wave Bye-Bye10. I am hungry

State of nature Type of gesture (1 … 10)

We assume that there is some a priori probability (i.e. prior) P(1) that the next gesture is ‘Big Circle’, P(2) that the next gesture is ‘Small Circle’, etc.If the gesture lexicon is finite:

Page 34: Bayesian  Decision Theory

Institute of Systems and RoboticsISR – Coimbra

Mobile Robotics Lab

Bayesian Approaches Bayesian Approaches 3434

jrettjrett

Missing and noisy features

Missing Features:

Example: x1 is missingmeasured value of x2 is x^2mean x1 points to omega 3but omega2 better decision