Project Report - University of California, San Diego

Download Report

Transcript Project Report - University of California, San Diego

Project Report
Anup Doshi
Online Learning, CSE290
Mar.20.06
“There is something I don't know that I
am supposed to know. I don't know what
it is I don't know, and yet am supposed to
know, And I feel I look stupid if I seem
both NOT to know it and not know WHAT
it is I don't know. Therefore, I pretend I
know it. This is nerve-wracking since I
don't know what I must pretend to know.
Therefore, I pretend I know everything.”
-R.D. Laing, Knots (1970)
A Mind-reading Race Game [demo]
General Outline
 A History: Shannon and Hagelbarger
 Motivation: Freund and Schapire
 Algorithms
 Data Analysis
 Future work
History
 Information Theory
Online Prediction
Universal Prediction
Data Compression
Lempel-Ziv, etc
 On a tangent:
Can we predict
humans?
Odd/even Games
History
Claude Shannon & D.W. Hagelbarger
Inspired by Edgar Allen Poe
“The Purloined Letter”
Describes a strategy to win an odd-even game
At Bell Labs in 1950s
Motivated by telephone systems
-Hagelbarger
History – Shannon vs. Hagelbarger
SEER, a SEquence Extrapolating
Robot
A Mind-reading(?) Machine
Finite State Machines
History – Shannon vs. Hagelbarger
More Recently…
Freund and Schapire
Interested in
Online Prediction
Human-Computer Interaction
Formulated Mind-reader as a racing game
Look at various algorithms
Hagelbarger
Dutch Trees (Context-Tree Weighting Method)
Combining experts (How to Use Expert
Advice)
Freund and Schapire’s Mindreader
Dutch Tree – Context Algorithm
Bounded Memory
Efficient Updating
‘Optimal’ Performance
Rissanen Lower Bound
Weighted Context Tree –
Extension of Variable-Length Markov
Models
See Talk 6 by Prof. Freund
Other Experts
Use Dutch Trees to predict a change
in user’s input
Shannon’s machine
Hagelbarger’s machine
Sleeping experts (for counting
sequence, etc.)
Could combine these experts as in
“How to use expert advice” paper
Data Analysis  plots…
Mindreading Game Data Analysis
number of games played over time
1200
1000
800
600
400
200
0
0
Mar.13.06
1
2
3
4
Day
5
6
7
Mindreading Game Data Analysis
number of games played over time
1200
1000
800
600
yoav emails entire CS dept
400
yoav emails COSMAL
200
0
0
Mar.13.06
1
2
3
4
Day
5
6
7
Mindreading Game Data Analysis
Histogram of Scores
80
70
60
50
40
30
20
10
0
-100
-50
Losers:
950/1190 (79.83%)
0
50
Winners:
240/1190 (20.17%)
Mindreading Game Data Analysis
Number of names used by single users
300
250
200
150
100
50
0
1
2
3
4
5
#names used
6
7
8
9
10
385 unique users (by IP)
Mindreading Game Data Analysis
Number of times played by single users
250
200
150
100
50
0
0
5
10
15
20
#times played
25
30
35
40
Mindreading Game Data Analysis
fraction won vs. games played
1
0.9
0.8
fraction of games won
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
5
10
15
20
25
number of games played by single users
30
35
40
Mindreading Game Data Analysis
fraction won vs. games played
1
0.9
0.8
fraction of games won
0.7
0.6
0.5
0.4
0.3
10 games won
9 won
0.2
6 won
5 won
0.1
1 won
0
0
5
10
15
20
25
number of games played by single users
30
35
40
Mindreading Game Data Analysis
Time taken vs. Score
4
3.5
Score: 41
Time: 11min 7.95sec
log10(seconds taken)
3
2.5
2
1.5
1
0.5
-100
-50
0
score
50
Mindreading Game Data Analysis
Time taken vs. Score
4
3.5
log10(seconds taken)
3
2.5
2
1.5
1
0.5
-100
-50
0
score
50
Mindreading Game Data Analysis
High Score: 43…fair and square
 I:00101110000101100011010000111010
1111101011011111110110001000110010
1100001100000111111101100100001100
1101010010101010111011110110011011
1111001111110101111100
 P:11000001101010110110001000101000
0011110100000000010111100111011100
1010110011010010010110110010000110
0110101001010101100100011001000010
0010001100101001011110
Mindreading Game Data Analysis
Low Score: -100
I:0101010101010101010101010101
01010101010101010101010101010
10101010101010101010101010101
0101010101010
P:0101010101010101010101010101
01010101010101010101010101010
10101010101010101010101010101
0101010101010
Mindreading Game Data Analysis
Counting sequence
Future work
Different game layouts
Timer to force a guess sooner
Sleeping experts & Sleeping experts
for specific strategies (e.g. counting)
Combining experts
Improve experts based on collected
data
Log-loss game (Cover’s horse-racing)
Cover’s horse racing – log loss
Cover’s horse racing
0
0.3
1
0.7