intelligent environments

38
Intelligent Environments 1 Intelligent Environments Computer Science and Engineering University of Texas at Arlington

Upload: purity

Post on 04-Jan-2016

52 views

Category:

Documents


0 download

DESCRIPTION

Intelligent Environments. Computer Science and Engineering University of Texas at Arlington. Prediction for Intelligent Environments. Motivation Techniques Issues. Motivation. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Intelligent Environments

Intelligent Environments 1

Intelligent Environments

Computer Science and Engineering

University of Texas at Arlington

Page 2: Intelligent Environments

Intelligent Environments 2

Prediction forIntelligent Environments Motivation Techniques Issues

Page 3: Intelligent Environments

Intelligent Environments 3

Motivation An intelligent environment

acquires and applies knowledge about you and your surroundings in order to improve your experience. “acquires” prediction “applies” decision making

Page 4: Intelligent Environments

Intelligent Environments 4

What to Predict Inhabitant behavior

Location Task Action

Environment behavior Modeling devices Interactions

Page 5: Intelligent Environments

Intelligent Environments 5

Example Where will Bob go next? Locationt+1 = f(…) Independent variables

Locationt, Locationt-1, … Time, date, day of the week Sensor data Context

Bob’s task

Page 6: Intelligent Environments

Intelligent Environments 6

Example (cont.)Time Date Day Locationt Locationt+1

0630 02/25 Monday Bedroom Bathroom

0700 02/25 Monday Bathroom Kitchen

0730 02/25 Monday Kitchen Garage

1730 02/25 Monday Garage Kitchen

1800 02/25 Monday Kitchen Bedroom

1810 02/25 Monday Bedroom Living room

2200 02/25 Monday Living room

Bathroom

2210 02/25 Monday Bathroom Bedroom

0630 02/26 Tuesday Bedroom Bathroom

Page 7: Intelligent Environments

Intelligent Environments 7

Example Learned pattern

If Day = Monday…Friday& Time > 0600& Time < 0700& Locationt = Bedroom

Then Locationt+1 = Bathroom

Page 8: Intelligent Environments

Intelligent Environments 8

Prediction Techniques Regression Neural network Nearest neighbor Bayesian classifier Decision tree induction Others

Page 9: Intelligent Environments

Intelligent Environments 9

Linear Regression

n

ii

n

iii

xx

yyxxm

1

2

1

)(

))((

bmxy

xmyb

x y

1 3

2 5

3 7

4 9

Page 10: Intelligent Environments

Intelligent Environments 10

Multiple Regression

n independent variables Find bi

System of n equations and n unknowns

nnxbxbxbby ...22110

Page 11: Intelligent Environments

Intelligent Environments 11

Regression Pros

Fast, analytical solution Confidence intervals

y = a ± b with C% confidence Piecewise linear and nonlinear regression

Cons Must choose model beforehand

Linear, quadratic, … Numeric variables

Page 12: Intelligent Environments

Intelligent Environments 12

Neural Networks

Page 13: Intelligent Environments

Intelligent Environments 13

Neural Networks 10-105 synapses per neuron Synapses propagate

electrochemical signals Number, placement and strength

of connections changes over time (learning?)

Massively parallel

Page 14: Intelligent Environments

Intelligent Environments 14

Computer vs. Human BrainComputer Human Brain

Computational units

1 CPU, 108 gates 1011 neurons

Storage units 1010 bits RAM,1012 bits disk

1011 neurons,1014 synapses

Cycle time 10-9 sec 10-3 sec

Bandwidth 109 bits/sec 1014 bits/sec

Neuron updates / sec

106 1014

Page 15: Intelligent Environments

Intelligent Environments 15

Computer vs. Human Brain

“The Age of Spiritual Machines,” Kurzweil.

Page 16: Intelligent Environments

Intelligent Environments 16

Artificial Neuron

Page 17: Intelligent Environments

Intelligent Environments 17

Artificial Neuron Activation functions

Page 18: Intelligent Environments

Intelligent Environments 18

Perceptron

Page 19: Intelligent Environments

Intelligent Environments 19

Perceptron Learning

Page 20: Intelligent Environments

Intelligent Environments 20

Perceptron Learns only linearly-separable

functions

Page 21: Intelligent Environments

Intelligent Environments 21

Sigmoid Unit

Page 22: Intelligent Environments

Intelligent Environments 22

Multilayer Network ofSigmoid Units

Page 23: Intelligent Environments

Intelligent Environments 23

Error Back-Propagation Errors at output layer propagated

back to hidden layers Error proportional to link weights

and activation Gradient descent in weight space

Page 24: Intelligent Environments

Intelligent Environments 24

NN for Face Recognition

90% accurate learning head pose for 20 different people.

Page 25: Intelligent Environments

Intelligent Environments 25

Neural Networks Pros

General purpose learner Fast prediction

Cons Best for numeric inputs Slow training Local optima

Page 26: Intelligent Environments

Intelligent Environments 26

Nearest Neighbor Just store training data (xi,f(xi)) Given query xq, estimate using

nearest neighbor xk: f(xq) = f(xk) k nearest neighbor

Given query xq, estimate using majority (mean) of k nearest neighbors

Page 27: Intelligent Environments

Intelligent Environments 27

Nearest Neighbor

Page 28: Intelligent Environments

Intelligent Environments 28

Nearest Neighbor Pros

Fast training Complex target functions No loss of information

Cons Slow at query time Easily fooled by irrelevant attributes

Page 29: Intelligent Environments

Intelligent Environments 29

Bayes Classifier Recall Bob example

D = training data h = sample rule

)(

)()|()|(

DP

hPhDPDhP

)|(maxarg DhPhHh

best

Page 30: Intelligent Environments

Intelligent Environments 30

Naive Bayes Classifier

Naive Bayes assumption

Naive Bayes classifier

),...,,(

)()|,...,,(maxarg

),...,,|(maxarg

21

21

21

n

jjn

Yybest

njYy

best

aaaP

yPyaaaPy

aaayPy

j

j

i

jijn yaPyaaaP )|()|,...,,( 21

ijij

YyNB yaPyPy

j

)|()(maxarg

y representsBob’s location

Page 31: Intelligent Environments

Intelligent Environments 31

Bayes Classifier Pros

Optimal Discrete or numeric attribute values Naive Bayes easy to compute

Cons Bayes classifier computationally

intractable Naive Bayes assumption usually violated

Page 32: Intelligent Environments

Intelligent Environments 32

Decision Tree Induction

Day

Time > 0600

Locationt

Time < 0700

Bathroom

M…F

yes

yes

Bedroom …

no

no

SatSun

Page 33: Intelligent Environments

Intelligent Environments 33

Decision Tree Induction Algorithm (main loop)

1. A = best attribute for next node2. Assign A as attribute for node3. For each value of A, create

descendant node4. Sort training examples to descendants5. If training examples perfectly

classified, then Stop, else iterate over descendants

Page 34: Intelligent Environments

Intelligent Environments 34

Decision Tree Induction Best attribute Based on information-theoretic

concept of entropy Choose attribute reducing entropy

(~uncertainty) from parent to descendant nodes

A1 A2

Bathroom (0)Kitchen (50)

Bathroom (50)Kitchen (0)

Bathroom (25)Kitchen (25)

Bathroom (25)Kitchen (25)? ? B K

v2v1 v1 v2

Page 35: Intelligent Environments

Intelligent Environments 35

Decision Tree Induction Pros

Understandable rules Fast learning and prediction

Cons Replication problem Limited rule representation

Page 36: Intelligent Environments

Intelligent Environments 36

Other Prediction Methods Hidden Markov models Radial basis functions Support vector machines Genetic algorithms Relational learning

Page 37: Intelligent Environments

Intelligent Environments 37

Prediction Issues Representation of data and

patterns Relevance of data Sensor fusion Amount of data

Page 38: Intelligent Environments

Intelligent Environments 38

Prediction Issues Evaluation

Accuracy False positives vs. false negatives

Concept drift Time-series prediction Distributed learning