zhiphone: a mobile phone that learns context and user preferences jing michelle liu raphael hoffmann...
TRANSCRIPT
Zhiphone: A Mobile Phone that Learns Context and User Preferences
Jing Michelle Liu Raphael Hoffmann
CSE567 class project, AUT/05
CSE567 project 212/8/2005
CSE567 project 312/8/2005
Objective
Adapt the alarm type of a mobile phone to its context.
CSE567 project 412/8/2005
CSE567 project 512/8/2005
Objective
Adapt the alarm type of a mobile phone to its context and the preferences of its user.
CSE567 project 612/8/2005
This is what you want
Track context
Receive call
User reaction
Learn user preferences
CSE567 project 712/8/2005
Outline
Related Work Architectural Overview Feature Extraction Learning Context Learning User Preferences Future Work
CSE567 project 812/8/2005
Related Work
SenSay: A Context-aware mobile phone Four states: uninterruptible, idle, active and
normal Five functional modules: the sensor box,
sensor module, decision module, action module and phone module
Sensors include light, motion and microphone Query sensor data and the electronic calendar
of the user Use thresholds to classify into different states
There is no learning from user preferences
CSE567 project 912/8/2005
Related Work
SenSay: A Context-aware mobile phone Four states: uninterruptible, idle, active and
normal Five functional modules: the sensor box, sensor
module, decision module, action module and phone module
Sensors include light, motion and microphone Query sensor data and the electronic calendar of
the user Use thresholds to classify into different states
There is no learning from user preferences
CSE567 project 1012/8/2005
Feature Extraction
Architectural Overview
0
111
Supervised Learning
Learning context
Reinforcement Learning
Learning user preferences
ringing tone vibration
offline
online
voice mail
raw data
features
soft sensors
alarm type
close to body conversation extr. noise outdoor physical act.
FFT1 variance1 variance2 derivative3
CSE567 project 1112/8/2005
Feature Extraction
potentially useful features running mean running variance derivative exponential smoothing frequency components (FFT)
short (10 sec) and long (1 min) running windows010111011010010011010100011110101010111010010111010100011010101
CSE567 project 1212/8/2005
Feature Extraction (cont’d)
system of pluggable components(using observer/composition design patterns)
rawsensor
windowfilter
windowfilter
FFTfilter
meanfilter
var.filter
windowfilter
meanfilter
joinfilter
CSE567 project 1312/8/2005
Feature Extraction - Experiments
measured execution times on the iPAQ
Feature Extractor Time
Mean (window 20) 0-1 ms
Variance (window 5500) 15-18 ms
Derivative 0-1 ms
Exponential Smoothing 0-1 ms
Fast Fourier Transform (window 4096)
1032-1083 ms
Window (window 5500, w/ overlap)
102 ms
CSE567 project 1412/8/2005
Feature Extraction
Architectural Overview
0
111
Reinforcement Learning
Learning user preferences
ringing tone vibration
online
voice mail
raw data
features
alarm type
Supervised Learning
Learning context offline
soft sensorsclose to body conversation extr. noise outdoor physical act.
FFT1 variance1 variance2 derivative3
CSE567 project 1512/8/2005
Learning Context
identified five relevant soft sensors Close to Body Conversation Extreme Noise Outdoor Physical Activity
Using annotated user traces, we build a Support Vector Machine classifier for each soft sensor offline
CSE567 project 1612/8/2005
Learning Context (cont’d)
Support Vector Machines
++++
++ +
++
+ +
+
++
-
-
--
--
-
Fea
ture
2
Feature1
? ??
CSE567 project 1712/8/2005
Learning Context (cont’d)
Support Vector Machines
++++
++ +
++
+ +
+
++
-
-
--
--
-
Fea
ture
2
Feature1
Maximize distance between closest
point and boundary(Optimization)
CSE567 project 1812/8/2005
Learning Context (cont’d)
Support Vector Machines w/ Gaussian Kernel
++
++
++ +
++
+ +
+
+
+
-
-
-
-
- -
-
Fea
ture
2
Feature1
-
CSE567 project 1912/8/2005
Learning Context – Experiments
recorded and annotated traces at Allen Center
Annotation Trace1 Trace2 Total
Close To Body 48 min 31 min 79 min
Conversation 24 min 5 min 29 min
Extreme Noise 4 min 9 min 13 min
Outdoor 6 min 10 min 16 min
Physical Activity
5 min 5 min 10 min
Total recorded time
62 min 55 min 117 min
CSE567 project 2012/8/2005
Learning Context – Experiments (cont’d) 2-fold Cross-Validation results (in %)
Annotation Precision
Recall Accuracy
Close To Body 88.058 87.647 82.799
Conversation 77.879 32.247 81.678
Extreme Noise 81.569 88.547 96.511
Outdoor 68.478 61.204 91.874
Physical Activity
94.309 79.293 97.876
Overall 82.059 69.788 90.148
CSE567 project 2112/8/2005
Learning Context – Experiments (cont’d) measured execution times on iPAQ and
desktopAnnotation Classificatio
nTraining
Close To Body 127 ms 4 sec (±2)
Conversation memory error
212 sec (±134)
Extreme Noise 74 ms 3 sec (±2)
Outdoor 132 ms 4 sec (±2)
Physical Activity
12 ms 3 sec (±2)
Overall >86 ms 45 sec (±28)
offlineonline
CSE567 project 2212/8/2005
CSE567 project 2312/8/2005
Feature Extraction
Architectural Overview
0
111
Supervised Learning
Learning context offline
raw data
features
soft sensors
Reinforcement Learning
Learning user preferences
ringing tone vibration
online
voice mail alarm type
close to body conversation extr. noise outdoor physical act.
FFT1 variance1 variance2 derivative3
CSE567 project 2412/8/2005
Learning User Preferences
Scenario 1 Soft sensors detect CloseToBody,
Conversation Call comes in and phone is ringing User hangs up phone without taking call
Scenario 2 Soft sensors detect CloseToBody,
Physical Activity Call comes in and phone vibrates User does not notice call
bignegativereward
smallnegativereward
CSE567 project 2512/8/2005
Learning User Preferences
Reinforcement Learning (Q-Learning)
Neural Network
soft sensor states+ current alarm type
new alarm type
CSE567 project 2612/8/2005
What we can do in the future
use an open-source VoIP library, e.g. JVOIPLIB, to enable real cell phone capability
apply more advanced reinforcement learning on user preferences
evaluate the effectiveness of learning user preferences
CSE567 project 2712/8/2005
End