playing gwap with strategies - using esp as an example
DESCRIPTION
Playing GWAP with strategies - using ESP as an example. Wen-Yuan Zhu CSIE, NTNU. Motivation. “Games with a purpose” (GWAP) is an innovative concept in computer science There are a lot of GWAP systems has been created Research on enhancing GWAP systems is scarce. State of the art. - PowerPoint PPT PresentationTRANSCRIPT
Playing GWAP with strategies- using ESP as an example
Wen-Yuan ZhuCSIE, NTNU
Motivation
• “Games with a purpose” (GWAP) is an innovative concept in computer science
• There are a lot of GWAP systems has been created
• Research on enhancing GWAP systems is scarce
State of the art
• “Human Computation” represents a new paradigm of applications– Some problems solved by human, not
computer• “Games with a purpose” (GWAP)
– created by Dr. Luis von Ahn (CMU)– the most popular application
Our Contribution
• We propose an evaluation metric for GWAP systems
• We study the inner properties of the ESP game using analysis
• We propose an “Optimal Puzzle Selection Algorithm” (OPSA)
Our Contribution (2)
• We implement a quasi ESP game, ESP Lite, to demonstrate our proposed algorithms
• We confirm GWAP systems are more efficient if they are designed and played with strategies
Outline
• Introduction• Analysis• Propose our algorithm• Evaluation & Simulation• Implementation• Result• Conclusion & Future work
Human Computation
• There is a lot of things that human can easy do that computers can not yet do– Speech recognition– Natural language understanding– Computer graphics
Games with a Purpose
• It combine computation with game • People spend a lot of time playing games• It makes Human Computation more efficient• There are a lot of GWAP systems has been
created (e.g. ESP game and Google Image Labeler)
What is the ESP game?Alice Bob
shoe flower
flowerrocks
agreement on “flower”
What is the ESP game? (2)
• it is efficient– 200,000+ players have contributed 50+
million labels– each player plays for a total of 91 minutes– 233 labels/human/hour (i.e. one label every
15 seconds)• Google bought a license to create its own
version of the game in 2006
Outline
• Introduction• Analysis• Propose our algorithm• Evaluation & Simulation• Implementation• Result• Conclusion & Future work
Objectives
1. Design a metric to measure the performing of the ESP game
2. Design proper strategies to improve the ESP game
3. Validate the proposed strategies in real world systems
Observations
• The ESP game has two goals1. Quantity : the system prefers to maximize
the number of images which have been played
2. Quality : the system prefers to take as many labels as possible for each image
• There is a trade-off between the two goals
System metricqualityquantityG
*)ln()ln( SNG
)ln()ln( rSr
TG
))ln()))(ln(ln()(ln( rSrTG
)ln()ln()ln())ln()(ln())(ln( 2 STrSTrG
CST
rG
2)2
)ln()ln()(ln(
n
iiSn
S1
1
i
iSNS
1*
average scores per labeled image# of labeled images
# of labels per labeled image
# of labels
Modeling(2)
2
)ln()ln( ST
er
2)2
)ln()ln((
STC
Outline
• Introduction• Analysis• Propose our algorithm• Evaluation & Simulation• Implementation• Result• Conclusion & Future work
Puzzle Selection Algorithms
• Optimal Puzzle Selection Algorithm (OPSA)– select an image based on our analysis
• Random Puzzle Selection Algorithm (RPSA)– select a image by random
• Fresh-first Puzzle Selection Algorithm (FPSA)– select a image that has been played least
frequently
OPSA
• The idea is there are optimal r labels per labeled image in system
• We group images into 3 sets– contains all the images that have not
been played– contains all the images that have been
played at least once, but less than r rounds– 102 PPPP
1P
0P
OPSA (2)
Outline
• Introduction• Analysis• Propose our algorithm• Evaluation & Simulation• Implementation• Result• Conclusion & Future work
Simulation
• Setup– There are 100,000 images– Running 20 times simulation
• Observation– T vs. r → OPSA– T vs. System gain → 3 strategies
• Discussion
T vs. r
2
])[ln()ln( SET
er
T vs. System Gain
Discussion
• OPSA is superior to RPSA & FPSA in the simulation
• A systematically & thorough study to verify the purposed strategies in real systems is highly desirable
• To this end, we decide to implement the ESP Lite system
Outline
• Introduction• Analysis• Propose our algorithm• Evaluation & Simulation• Implementation• Result• Conclusion & Future work
ESP Lite
• We implement ESP using Flash/Java• It is a mimic ESP game• It implemented three playing strategies
– RPSA, FPSA, OPSA
Flow chart
the strategy selection process gives priority to the strategy that has been used least in terms of the # of rounds played previously
Client interface
Client interface (2)
Score System
• The score system is used to measure the quality of the agreed words
• The quality of each the agreed word should depends on its popularity– high frequency → low quality → low score– low frequency → high quality → high score
Score System (2)
i-th level of the w_i
nwww ,,, 21 n words in the score table
nfff 21 frequency of the word
0, 0
0
1
0
fkf
f
L n
jj
i
jj
i
baseoffsetii SSLwscore )( score of w_i
4
1 levels score
levels score 5
k
reserved for agreed words which are not in score table
Score System (3)
• Apply the Porter Stemming Algorithm to remove common morphological and inflectional endings of English words– Prevent words with the same root, but
receiving different scores (e.g. determinant and determine)
– Reduce the plural form to the singular form (e.g. experiments and experiment)
Outline
• Introduction• Analysis• Propose our algorithm• Evaluation & Simulation• Implementation• Result• Conclusion & Future work
• 100,000 images from ESP dataset• ESP dataset
– 100,000 images– Average 15 labels per image collected from
the ESP game• Score system
– – Range of score (10 levels)
Experiment Setting
9k
}150,,70,60{
10scaleS60baseS
Experiment Setting (2)
• Score system (cont.)– The 5,000 most frequency words from
Brown Corpus (a general corpus in the field of corpus linguistics)
– Processed by Porter Stemming Algorithm – 3,476 words in score table
Experiment
• http://nrl.iis.sinica.edu.tw/GWAP/ESPLite/• From 2009/3/9 to 2009/4/9• 3,103 games• 9,376 labeled pictures• 12,312 agreements
Score Statistics
Behavior of OPSA
• What happened in each round?
Behavior of OPSA (2)
Behavior of OPSA (2)
• The operation of OPSA depend on r
• Observe players’ behavior
Behavior of OPSA (3)
Performance
Outline
• Introduction• Analysis• Propose our algorithm• Evaluation & Simulation• Implementation• Result• Conclusion & Future work
Conclusion
• In this thesis– We propose a metric to evaluate the performance
of GWAP– We play GWAP systems with strategies and make
GWAP systems more efficient
Conclusion (2)
• This is the first GWAP study that implements and evaluates an analytical model on real-world GWAP systems
• our experiment results confirm that GWAP systems are more efficient if they are designed and played with strategies
Future work
• Thinking about the time factor in analysis• Thinking about adaptive r for each image• Constructing the framework of GWAP systems• Developing the GWAP portal for fast accessing
more GWAP systems• Developing the GWAP systems toward mobile
environment
Thank You!