萬能科技大 學資工系 助理教授 徐旺興 Email: [email protected]

萬能科技大學資工系 助理教授 徐旺興Email: [email protected]

Computational Intelligence and its Applications計算智慧及其應用

OutlineMotivationComputational Intelligence

◦Fuzzy, ANFIS, SVM, HMM and GMMFrequency Calibration based on the

ANFISHandwriting Recognition on Handheld

Devices using AccelerometersConclusionFuture work

Motivation (1/2)Studied in past years

◦QoS: Simulation◦NGN: Framework improvement◦SIP: multimedia (client/server)◦Handoff: BS, Agent or Broker and

Mobile Device◦Time & Frequency: Control system◦3D Handwriting recognition

Motivation (2/2)Issue of time series

What is Computational IntelligenceCI related to other branches of

computer science, such as artificial intelligence (AI), classification, data mining, graphical methods, intelligent agents and intelligent systems, machine intelligence, machine learning, natural computing, parallel distributed processing, pattern recognition, probabilistic methods, soft computing, multivariate statistics, optimization and operation research.

Frequency Calibration based on the ANFIS

Implement a control systemNormal mode and holdover modeTechnology of CI

Fuzzy - a mathematical system that analyzes analog input values in terms of logical variables that take on continuous values in interval [0,1], in contrast to classical or digital logic, which operates on discrete values of either 0 or 1 (true or false).

ANFIS - learns features in the data set and adjusts the system parameters according to a given error criterion.

The requirements of time and frequency accuracy for the dominant wirelesstechnologies.

OCXO (Oven-Controlled Crystal Oscillator)

The stability of OCXO base on environmental effects such as vibration, temperature, pressure and humidity.

System architecture (1/2)

Voltage: +/-10V

Slave Clock1pps1pps

1pps=one plus per second

Frequencyoffset

System architecture (2/2)The frequency offset with respect to time Their change are the input variables of the fuzzy controller

)( ity

)()()( 1 iii tytyty

An incremental voltage generated by the fuzzy controller is used to update the voltage for steering the oscillator below.

)( itV

)(1 iii tVVV

Fuzzy rule table

NBNBNBNBNBPB

NBNSNSZEZEPS

NSNSZEPSPSZE

ZEZEPSPSPBNS

PBPBPBPBPBNB

PBPSZENSNB

NBNBNBNBNBPB

NBNSNSZEZEPS

NSNSZEPSPSZE

ZEZEPSPSPBNS

PBPBPBPBPBNB

PBPSZENSNBy

y

Ri : if y is Ai1 and y is Ai2 then u is Bi, for i 1,2…n

Ai1

Ai2

Bi

The input space is divided into five sets: negative big (NB), negative small (NS), zero (ZE), positive small (PS) and positive big (PB) for a frequency offset or its change.

Component of the system (1/2)Cesium (HP5071A)

◦10 MHz of a cesium atomic clock (10-14)

OCXO (FTS1130)◦10 MHz of oven-controlled crystal

oscillator (10-8)TIC (SR620)

◦time interval counter◦Time interval and frequency counter

Component of the system (1/2)D/A

◦ADLINK PCI-6208◦16-bit resolution with the bi-polar ◦Voltage: 10V to +10V

Fuzzy controller◦Software coding by C/C++, Matlab

ANFIS controller◦Software coding by C/C++, Matlab

Two mode in this systemNormal Mode

◦Fuzzy controller◦Collecting the control signal◦To train the ANFIS controller

simultaneouslyHandover Mode

◦ANFIS controller◦When the signal of the primary clock

is lost◦The voltage (control signal) is

predicted by the ANFIS controller

Experimental setupThe OCXO is steered every 10s

by the fuzzy controller or the ANFIS controller to syntonize with the primary clock.

Choice about five-day input-output data pairs:

The first four-day pairs were used for training the ANFIS.

The remaining about one-day pairs were used for validating the identified model.

)]();(),(),2(),3([ txtxtxtxtx

Experimental results and analysis

The desired data

The predicted

data

The Prediction

error

Frequency stability

100

101

102

103

104

105

10-14

10-13

10-12

10-11

10-10

10-9

Frequency stabilityM

odifi

ed A

llan

devi

atio

n

Normal

Holdover

Free Running

Averaging time (seconds), Averaging time (seconds),

)( yMod

Conclusion of this chapterThe frequency stability of the

OCXO could be improved from a few parts in 10-9 to 10-12 over a measurement period of one day. (Normal Mode)

Holdover Mode shows the frequency stability of the OCXO could be maintained within a few parts in 10-11 for an averaging time of on day.

3D Handwriting RecognitionAccelerometer3D gesturePattern RecognitionMobile Device’s Accelerometer

(30Hz~70Hz)Device

◦HTC G1Software component

◦Collectors (Java code)◦Training and recognition (matlab code), off-line.

Proposed methodWLCS + SVMHMM + GMM

The architecture of the proposed 3D handwriting recognition system

Three axes acceleration data of pattern ‘Kimble’.

Data preprocessing

Data Training

Longest Common Subsequence （ LCS ）Example of LCSs1: 2 5 7 9 3 1 2s2: 3 5 3 2 8

LCS: 5 3 2

Weight LCS

Data Classification – SVM (1/3)

To Find the Hyper-plane (e.g. 2D’s hyper-plane is a line)

SVM (2/3)

To Find the optimal Hyper-plan H

SVM (3/3)

Experimental setupWe collect a set of 26 gestures

(alphabet), 20 samples per gesture from 3 different persons, totaling 1560 gestures samples.

50 samples for training and 10 samples for testing.

The average length of the WLCS between letters

Models

Test data

Performance CriteriaThe classification performance

can be evaluated using mis-classification rate such as apparent error rate and/or graphical representation tools such as the receiver operating characteristic (ROC) curve.

Terms associated ROC curve

An example of ROC

The table shows 20 data and the score assigned to each by a scoring classifier

Sorting by score

ROC curve of the example

10 positive points at

x-axis

10 positive points at y-

axis

Max. and Min. AUC (Area under curve)

The Max. AUC is

alphabet’C’.

The Min. AUC is

alphabet ’G’

List of AUC from ‘A’ to ‘Z’the alphabet

such ‘C’, ‘L’, ‘P’, ‘S’, ‘U’, ‘V’ and ‘Z’ is good instance and

’G’ is a randomly chosen negative instance.

Summary about LCS +SVMLCS + SVM is the lite-computing

algorithm, the average accuracy is 86.85%

Hidden Markov Model (HMM)HMMs allow you to estimate

probabilities of unobserved events.

Given plain text, which underlying parameters generated the surface.

E.g., in speech recognition, the observed data is the acoustic signal and the words are the hidden parameters.

HMMs and their UsageHMMs are very common in

Computational Linguistics:◦ Speech recognition (observed: acoustic

signal, hidden: words)◦ Handwriting recognition (observed: image,

hidden: words)

Parameters of an HMMStates: A set of states S=s1,…,snTransition probabilities: A= a1,1,a1,2,

…,an,n Each ai,j represents the probability of transitioning from state si to sj.

Emission probabilities: A set B of functions of the form bi(ot) which is the probability of observation ot being emitted by si

Initial state distribution: is the probability that si is a start state

i

The Three Basic HMM Problems (1/2)Problem 1 (Evaluation): Given the

observation sequence O=o1,…,oT and an HMM model

, how do we compute the probability of O given the model?

Problem 2 (Decoding): Given the observation sequence O=o1,…,oT and an HMM model

, how do we find the state sequence that best explains the observations?

(A,B, )

(A,B, )

Problem 3 (Learning): How do we adjust the model parameters , to maximize ?

The Three Basic HMM Problems (2/2)

(A,B, )

P(O | )

Example of HMM

The states

The observations

An Example model, the semi-code of HMM

The dynamic programming computation

Diary data and reconstructed weather

In this workGiven the observation sequence

O=o1,…,oT, ◦e.g.

5555222233344433377001111111….

Build 26 Model HHMA, HHMB, … HMMz

Training each model by EM algorithm. (Problem 3)

Recognition, compute the probability of O given the model. (Problem 1) (Forward-Backward Algorithm)

GMM – Gaussian Mixture model

Experimental setupWe collect a set of 26 gestures

(alphabet), 20 samples per gesture from 3 different persons, totaling 1560 gestures samples.

The list of log likelihood compare with “Model” and “Test Data”

Summary about HMM+GMMUsing GMM, the accuracy of

classification could be achieved at about 96.5%.

CONCLUSIONFrequency calibration:

◦This study demonstrates the feasibility of using the ANFIS for frequency calibration.

◦The frequency stability of the OCXO can be significantly improved more than three orders of magnitude in both normal mode and holdover mode.

3D handwriting gesture recognition: ◦This study demonstrates the feasibility of using◦method of (SVM and LCS), and using method of

(HMM and GMM) for 3D handwriting gesture recognition.

Future workNew Tools of Computational

Intelligence ◦HHT(Hilbert Huang Transform)◦http://rcada.ncu.edu.tw/

New Area about signal process◦PhysioNet

生醫訊號資料庫http://www.physionet.org/

PhysioNet包含：

◦PhysioNet ： the research resource for complex physiologic signals Publications Tutorials Challenges

◦PhysioBank ： database◦PhysioToolkit: software