Using High Performance Computing Nov 1 3 200 9 AESCS 200 9 Taiwan Takao TERANO Department of Computational Intelligencs amp Systems Sciences Tokyo Institute of Technology ID: 784459
Download The PPT/PDF document " Agent-Based Simulation in History Re..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Agent-Based Simulation in History Research Using High Performance Computing
Nov
., 1
3
, 200
9
AESCS
200
9
, Taiwan
Takao
TERANO
Department of Computational Intelligencs & Systems Sciences
Tokyo Institute of Technology
terano@dis.titech.ac.jp
www.trn.dis.titech.ac.jp
Slide2One Minute Summary Background:High Performance ComputingComputational Social Science
Motivation:
Investigate “Past Events” by Rewinding the tape
Topic We Concern:
Civil Servant Examination System (
科挙試験
) in China over 1,400years
Approach:
Data Analysis
(Mining!)
Agent-Based Approaches with Large-scale Optimization Algorithms & a Grid Environment
Result:
“Art is a lie that helps us see reality” by Pablo Picasso
Also , ABM with High Performance Computing is …
Slide3AgendaIntroduction Computational History Research Description of Chinese CS Exams
Modeling Genealogical Records to Agents
Optimization of ABS & GOGA Framework
Experimental Results & Findings
Concluding Remarks
Slide4TSUBAME Grid at Titech
No 9 from the Data “Top 500 inNov., 2006”
Slide5History Research & Archaeology- The only discipline with the time range to study long-term culture change.Agent based modeling allows us to “rewind the tape” of cultural evolution.Impossible to experiment & to make verification difficultExample: Artificial Anasazi
Slide6Brief Descriptions of Civil Servant Exams (科挙試験) in China
Establish and Keep the Bureaucracy to Support the Emperor of Chinese Dynasties
Origin
：
Han(
漢
)era
（
3c
），
Start:
Sui（隋）
era (6c), Continue: Tang(
唐
),
Establish:Song(
宋
)(10c), Thrive:Yuan(
元
)
、
Ming (
明
)
、
end:Ching
（清）
(20c)
The Longest History: over 1,400 years continued
Applicants: over 300k persons/yr; #Passed: 300 persons/3yrs (in Ching era)
3 or 4 levels Exams (Regional to Central)
Levels of Candidates:
sheng-yuan
, kun-sheng; chien-sheng; chu-jen; kung-shih; chin-chi(
進士
) (6 levels)
Exam Questions: Liberal Arts (Not Natural Sciences)
2-3 days exams!
They did not make educational systems but only exam. systems
Slide7Roles of women in Cultural Capital Transfer“A wise daughter will make a wise wife and mother”(Mann).The Marriage System in ChinaSince the sixteenth century, the upper class families in Hunan area in China focused on the levels of women’s cultural and literary talents are critical factors for successful marriage. (Ko
)
History of Civil Service Examination
It is an examination for recruitment of government officials, which has been run for about 1,300 years.
Brief Literature Review in Family Systems
Slide8The effect of relatives
Elman(1991)
It was possible for ordinary candidates to achieve academic success, because they had bureaucrats among their relatives.
To create elite families ensured success of the examinations.
It guaranteed that such the background gained successful future and political career.
The candidates came from a family which had the tradition of learning classics and the official language.
Certain family lines could produce more successful candidates.
Slide9Successful Family Y The Y family in Changzhou, Jiangsu are is a typical caseY produced twenty-two successful candidates for the period of more than twelve generations.
There are genealogical records (Zupu)of 1,237 members between 1468 and 1944 (China and Japan).
Zupu is a paternal (fathers
’
) record including name, birth year, year of death, achievement in life, wife
’
s name, number of children, place of residence and other information for each family member, which consists of mainly two parts:
“Shixi" : general family tree.
“Shibiao“: the details of the profile of each member.
Each example is shown in the next figures.
Slide10Family Network
Successful
Cluster
Unsuccessful Cluster
Slide11Shixi of the Y family
Slide12The Y family’s profile (“Shibiao”)
Slide13Research Questions:Why Successful and Un-SuccessfulsHow were they EducatedHow they Transmit the Knowledge and Roles of Family Systems
Norms, Marriage Systems, (In)tangibles,…
The Questions are hard to Answer only with Literature Survey Studies
ABM!
Slide14Zupu Pre-ProcessingThe adjacency matrix which is changed from the “shixi
”
: the family tree.
The attribute matrix which is changed from the
“
shibiao
”
: the family profile.
Shixi: The Adjacency Matrix
Shibiao: The Attribution Matrix
Slide15Outline of the Agent SimulationScenarios Each agent can transmit cultural capital,
from parent to child, from great-grandfather to great grandson, from grandfather to grandson, or from uncle and aunt to nephew.
There are two categories of cultural capital.
knowledge and art
Children have two cultural characteristics by birth.
knowledge and art
A child's cultural capital depends on the
synthetic
effect s of his characteristics and transmitted cultural capitals.
Only knowledge cultural capital affects success in the examinations.
Art cultural capital does not directly affect the rate of success.
Slide16Parameters:Who is the transmitter (great-grandfather, grandfather, father, mother, uncle, aunt).Degree of effect on individual cultural capital (rate of cultural transmission from father and others)Education effect by child’s cultural characteristics.
Degree of effect of the mother’s family home (transmission rate of cultural capital).
Degree of effect of the aunt’s married family (transmission rate of cultural capital).
Slide17Relation among K &A C.P.
Slide18Cultural capital transmission function
Slide19Inverse Simulation
Forward Simulation
Design the Model
Set Various Parameters
Execute Simulation
Evaluate Results
They consider the approach Very Difficult!!
GA techniques work well!!
Inverse Simulation Method
Design a Model with Many Params.
Set a Global Objective Fnc.
Execute Simulation to Optimize it
Evaluate Initial Parameters
Slide20Avoid manual parameter tuning
Evolve ‘good’ societies based on fitness functions associated with Socio-Metrics
Analyze the characteristics of the agents in the Evolved Society
Evaluation
Selection
Crossover
Mutation
Genes of Society
Simulation of Artificial Societies
Fitness=Macro-Level Socio-metrics
n-interval
n-interval
Micro-Level
Behavior
Pre-determined Features
Acquired Features
Micro-Level
Phenomena
Evolving Societies by GA:
Inverse Simulation
Slide21Assumptions for IS
Micro-Level Rich Functionality of the Agent Simulator with Enough Number of Parameters
Macro-Level Clear Specification of the Desired Results like
min f(…)
Fast Execution of the Simulation
Good GA-Based Techniques
Slide22History Simulator with IS
F
M
F
M
F
F
M
1
0
0
0
1
0
0
0
1
1
0
0
0
0
1
1
0
0
0
1
0
0
0
1
Family Tree
Rules, Parameters
Simulated
attribution
matrix
Real
attribution
matrix
1
0
0
0
1
0
0
0
1
1
0
0
1
0
0
0
1
0
0
0
1
1
0
0
GA
1
0
0
0
1
0
0
0
1
1
0
0
1
0
0
0
1
0
0
0
1
1
0
0
1
0
0
0
1
0
0
0
1
1
0
0
Evaluation
If A then X
If B then Y
If C then Z
P1=0.3
P2=0.7
MAS
If A then X
If B then Y
If C then Z
P1=0.3
P2=0.7
MAS
Agent
simulator
0
0
1
1
0
0
0
1
0
0
0
1
Evaluation
F
M
F
M
F
F
M
Slide23Experiments without Grid
if (maternal grandfather is a successful candidate) then
a child gets his cultural capital
else a child doesn’t get it
Effect by education = 0.4
Effect by cultural capital = 0.4
Crossover rate of cultural capital = 0.2
Obtained Strategies:
Slide24Who Educated Children?
Culture Transmitter to Children:
Grand Father
Generation Effects: Father : Granpa=1.0 : 2.8
F
M
F
M
F
F
M
F
M
F
M
F
F
M
F
M
F
M
F
F
M
Parents
Parents & Granpa
Granpa
Slide25What Let Children Learn?
Both Knowledge & Cultural Capitals from Parents(type5)
K Capitals of Child
= r*[KC of Parents]
*
[KA of Child]
+ (1-r)*[AC of Parents]
*[CA of Child] A Capitals of
子Child = r*[AC of Parents]*[CA of Child] + (1-r)*[KC of Parents]
*
[KA of Child]
Cultural Crossover Rate: 0.2
K
A
K
A
Balanced
0.8
0.2
0.8
0.2
K
A
K
A
K. Directed
Granpa, Parent
Child
1.0
1.0
Slide26Who made them successful, again?
Women’s Roles
if (Mothers’ Father Succeeded) then 1.0
otherwise, 0.0
F
M
F
M
F
M
F
F
M
F
Support from
Wife Giver
Education form Father to Daughter
Education
From Mother
To Child
Marriage
Slide27Desirable Three types of ABS
Forward Simulation
Design a
Model
Set Various Parameters
Execute Simulation
Evaluate Results
Design a
Model
Generate Parameters based on
Optimal Algorithm
Execute Simulation
Evaluate the Parameters
Inverse Simulation
Model Selection
Repeat
Design a Model
Generate a Subset of Parameters
based on Optimal Algorithm
Execute Inverse Simulation
based on this Subset of Parameters
Evaluate the Subset of Parameters
Optimize
Optimize
Slide28Forward Simulation Architecture
Slide29Inverse Simulation Architecture
[Yang et al. 2009]
Slide30Model Selection Architecture
Slide31Real-coded GA(UNDX+MGG) [Ono]Real-coded GA to search for appropriate parameters in the simulation.
Computationally Intensive! GOGA framework for parallel computation.
Slide32GOGA Framework
Slide33Objective Function of H. ABS
Slide34Experimental Setups and Results Configurations :
selection by tournament
crossovers of MGG:
200
alpha of UNDX:
0. 5
beta of UNDX:
0.35
number of societies:
50
number of generations: 2000.
Parameters
Results
People who transmit cultural capital to child
father, uncle, mother
, aunt
Education Effect by the characteristic of child
3%
Effect by mother
28
%,if
maternal
grandfather is a successful candidate.
Effect by aunt
6%, if father in aunt’s married family is a successful candidate.
Crossover rate from art cultural capital to knowledge cultural capital
100%
Crossover rate from knowledge cultural capital to art cultural capital
0
Transmission method of cultural capital
Both cultural capitals of knowledge and art are transmitted to the child from parents.
Slide35Cultural Capital Transmission Functions
There is a strong bond in
the exchanges between artists and intellectuals, and
the relationships between brothers and sisters
Slide36Strategies for Success!?
Father & Mother Necessary but not Sufficient
Families are important
Education through Granpa, Uncles, & Aunts
With Granpa & Uncles, Let children learn cultures
Artistic Education Necessary
Let children learn poem, music, …, & Friends!
Wife from/to Excellent Families
Learn good cultures from wives & aunts
Slide37Experimental Configuration and Results Configurations :
selection by tournament
crossovers of MGG:
200
alpha of UNDX:
0. 5
beta of UNDX:
0.35
number of societies:
50
number of generations: 2000.
Parameters
Results
People who transmit cultural capital to child
father, uncle, mother
, aunt
Education Effect by the characteristic of child
3%
Effect by mother
28
%,if
maternal
grandfather is a successful candidate.
Effect by aunt
6%, if father in aunt’s married family is a successful candidate.
Crossover rate from art cultural capital to knowledge cultural capital
100%
Crossover rate from knowledge cultural capital to art cultural capital
0
Transmission method of cultural capital
Both cultural capitals of knowledge and art are transmitted to the child from parents.
Slide38Model Comparison
Forward Simulation
Self Regression Model
Generalized Linear Model
MSE=6.23
MSE=1.12
MSE=1.92
MSE=4.75
Inverse Simulation
Slide39Roles of Male MembersFather and uncle play most important roles in the family for transmitting cultural capital to a child, and the influence of father is a little stronger than that of the uncle.
Great-grandfather has little influence on his great-grandson, which indicates that generation distance over three transmits little cultural capital to a child.
The important role of a grandfather to his grandson is replaced by the uncle‘s role ,which is because China has a big family system and all the family members live together, then uncle inherits the cultural capitals from grandfather and transmit them to the child indirectly.
Slide40Roles of Female Members The roles of mothers and aunts in a successful cluster are influential.
Influential effects of aunt are discovered [Ko] from case study:
Which coincides with the Ko’s case study:
Families, which belong to the same class and keep the relationship of marriage over generations, bring up their children each other.
Elite families to supplement marriage alliances, or simply to enhance a son's educational opportunities.
Influential effects by the aunts, which are relatively smaller than the ones by the mothers, although they would keep the norm of a reciprocal relationship between families.
Slide41Comparison: Winners & LosersFather: the Most Important Cultural Transmitter
The Cultural Influence to the Child:
Birth family of Mother > Married Family of Aunt.
Slide42Transmission of Knowledge
Slide43The Strategy to Success for ExamsFather has the same responsibility as mother to educate children.It is significant that combination effects of father, uncle, mother and aunt are to maintain a successful family norm.
Both mother in wife-taker and aunt in wife-giver have positive influence to success.
A child has more possibility to grow up to a successful candidates when he has more chances to appreciate art.
Education system is helpful to children.
Slide44Concluding Remarks
We have proposed a new method of Agent-Based Simulation (ABS) Model using a family tree to study history and cultural anthropology
Using a Grid System, Forward-, Inverse-Simulation, & Model Selection Mechanisms work well
We have analyzed successful candidates of Chinese civil service examinations with the ABS, which is based on the principle of cultural capitals [Bourdieu 1979]
We demonstrated that the roles of members of a family, especially artists are important to transfer the cultural capitals.
Slide45Two MessagesHigh Performance Computing is Critical for Cutting-Edges of both Social Science Researches and Design & Implementation of Social Architectures!
Agent-Based Modeling is a New Art, and
“
a lie that helps us see reality”
Slide46Acknowledgements・Contributors:
Prof. Setsuya Kurahashi(Tsukuba U.), Dr. Keiko Kurahashi(
Rikkyo
U.),
Prof. Isao Ono, Ms. Chao Yang, and Mr. Toru Takahashi (Titech)
・
References
[1] Setsuya Kurahashi, Takao Terano: Historical Simulation: A Study of Civil Service Examinations, The Family Line and Cultural Capital in China. Advances in Complex Systems (ACS) ,Vol.11, No. 2, pp. 187-198 (April 2008)
[2] Chao Yang, Setsuya Kurahashi, Keiko Kurahashi, Isao Ono and Takao Terano: Agent-Based Simulation on Women's Role in a Family Line on Civil Service
Exami
-nation in Chinese History. Journal of Artificial Societies and Social Simulation vol. 12, no. 2,2009 http://jasss.soc.surrey.ac.uk/12/2/5.html
[3] Chao Yang, Toru Takahashi, Bin Jiang, Isao Ono, Setsuya Kurahashi, Takao Terano: A Grid-Oriented Social Simulation Framework for
Large Scale Agent-Based Modeling, Proc. 6-th Conference of
European Social Simulation Association (ESSA 2009), 2009.