AN AGENT BASED ESTIMATION METHOD OF HOUSEHOLD MICRO-DATA INCLUDING HOUSING INFORMATION FOR THE BASE YEAR IN LAND-USE MICROSIMULATION

Transcription

1 AN AGENT BASED ESTIMATION METHOD OF HOUSEHOLD MICRO-DATA INCLUDING HOUSING INFORMATION FOR THE BASE YEAR IN LAND-USE MICROSIMULATION Kazuaki Miyamoto, Tokyo City University, Japan Nao Sugiki, Docon Co., Ltd., Japan Noriko Otani, Kansai University, Japan Varameth Vichiensan, Kasetsart University, Thailand 1

2 Introduction Land use microsimulation model - Requires micro-data for the base year including many attributes Micro-data not available - Synthetic population - Created from the accessible aggregated data Theoretically sound methodology - Synthesize household data 2

3 Objectives Develop a consistent method for estimating a set of agent based household micro-data - Extension of prior study (limited only to gender and age of each members) to comprehensive household data - Including housing type and the location Case Study for Validation - Apply the system to Parson-Trip-Survey data in the Sapporo Metropolitan area, Japan 3

4 Previous Works Iterative Proportional Fitting (IPF) based Methods Deming and Stephan (1940) - Population Synthesis Beckman et al. (1996) Miyamoto et al. (1986) - Improved IPF Guo & Bhat (2007) Pritchard & Miller (2009) Agent Based Approaches - Monte-Carlo Sampling by Agent Moeckel et al. (2003) Miyamoto et al. (1986) 4

5 Difficulties in Population Synthesis IPF with complex dataset - Zero-cell problem - Adding attributes reduce reliability Confidence or rigidity of classification - How much we can rely on - Similar to MAUP (Modifiable Areal Unit Problem) - Different result with different zoning 5

6 Presupposition The target micro-data of the households - Relationships of the members with the household head - Gender and age - Housing type and the location in the zone. Marginal condition from the census data - Number of households by the number of members - Number of persons by five-year age bands and gender - Number of housings by type - Number of households in each zone A certain number of samples that contains full information of the micro-data are available. 6

7 Micro-Data Dataset for household s with m members A ( hhl) { c, x, j, z} c ms ms ms ms : Member composition x ms : Age composition j : Type of housing z : Located Zone General form of Micro-data { ai 2 ( ai 1, ai,, air) 1 i N} R : Number of possible member types N : Number of households in the study area a ik : Age of member in household i Micro-Data ={head (male), head (female), husband, wife, one child (male), three children (male), one child (female), three children (female), father, mother, one grandchild (male) } ={45, 999, 999, 42, 15, 999, 999, 999, 999, 999, 999, } 7

8 Basic Estimation Prinsiples Estimation Based on Probabilities Obtained from Sampled Micro-dataset Relationship or contingency - Between attributes of household members Adjustment to satisfy the marginal condition - Number of persons by age band by gender - Number of housings by type by zone Probabilistic Process by Mote Carlo Approach - Synthesis - Adjustment 8

9 Correlation between Continuous Attributes Age of household members Original attribute variables : Non-correlated variables : 1 x p ( x, x2 p, p,, x,, p 1 m ( 1 2 m ) ) ran is Random number generator (A Random Number) Cumulative frequency curve of p i p Vx x V 1 p Wp 0 min Pis P is max P is 9

10 Flow of Estimation System ms START Y m M 1 s N m 0 m m 1 N s 0 s s 1 ( ms S m Y 1 random number generator ran s age composition of household x ms s [ x s, x2 x, x,, x,, x 1 s ms [ 1c 2c mc h ms { c }, xms Sample number (n) of m-member households n : 1,2,, N m C(n) [member composition of sample number (n): head(meber1) s gender, member2 s relation to head and gender, member3 s relation to head and gender,, member(m) s relation to head and gender] jms INT( Nm * rans ) 1 c C j ) ms ] ] micro data of household Initial set of synthetic households (at the initial calculation) s c Y s C rare N random number generator ran is i 1, m x is i 1, m population check by age band by gender t N y T y linear equations p Vx x Wp Y if x N ks xms belongs m to the age band by gender pis vik xks i 1, m k gy ; tgy T gy m xis wik pks i 1, m k age composition of household s x micro data of household s cumulative curve of p is i 1, m ms hms [ x s, x2 N,, x 1 s ms { cms, xms} Adjustment to marginal conditions population by age band ( y ) by gender ( g) t gy random sampling m, s h s if regenerated improves the difference between t and T gy gy for all gy ] ( in adjustment calculations) Y 1 Y 10

11 Flow of Estimation System Y m M 1 s N 1 m 0 m m 1 N s 0 s s 1 S m Y 1 random number generator ran s Initial housing type setting prob ( z) exp( V ) / exp( V ) s sz L l prob s(1) prob s(2) prob s (z) prob s (Z) 0 zone of household random number generator ran s prob s (z) is the probability of that s lives in zone z given by the logit model s sl ran s number of type housings in zone number of housings check by type by zone t N jz T jz j z random sampling m, s Y if (hhl) ms belongs to the housing type in zone ( j, z) ; prob s ( j) is the probability of that s lives t jz T jz in housing type j given by the logit model K prob s( j) exp( Vsj) / exp( Vsk) age composition of k household prob s (1) prob s (2 ) prob s ( j ) prob s (K ) N if regenerated ( j, z) xms [ x1 s, x2s,, xms ] improves the difference between t jz and T jz micro data of 0 ran 1 for all ( j, z) s household housing type of household s = j ( hhl) ms { cms, xms, j, z} ( at the initial calculation ) ( in adjustment calculations ) = z Initial living zone setting Adjustment to marginal conditions Y N end Y 11

12 Evaluation of Goodness-of-fit To develop the system in a more rational and objective manner Goodness-of-fit indicator between two of micro-datasets by Paper 2332 GOODNESS-OF-FIT EVALUATION METHOD BETWEEN TWO SETS OF HOUSEHOLD MICRO-DATA FOR LAND-USE MICROSIMULATION MODEL - Defined using the minimum value of the normalized sum of weighted distances - Solution algorithm using the genetic algorithm (GA), especially symbiotic evolution 12

13 Data Sapporo Metropolitan Area Person Trip Survey Totally 19,394 households data with full-scale information Random sampling 10,000 households data to compose a virtual dataset A Random sampling (11,367 males and 12,748 females) 1,000 households data to compose a sample dataset B Marginal Condition Data - Number of households with m members (m = 1, 2,, 7) - Number of individuals by gender and five-year age band - Number of households by type of housing by zone 13

14 Housing Types and Zone Setting 5 types of housing and 8 zones Housing types - Own-detached - Rent-detached - Own-apartment - Rent-apartment - Other Zone setting zone6 zone4 zone5 zone1 zone2 CITY CENTER zone3 zone7 zone8 Housing Type Samples of Dataset A Own-detached 5,233 Rent-detached 329 Own-apartment 1,395 Rent-apartment 2,889 Other 154 Total 10,000 Zone Samples of Dataset A zone 1 1,095 zone 2 1,972 zone 3 1,254 zone 4 1,940 zone 5 1,551 zone zone zone Total 10,000 14

15 Household Member Types and Membership Composition 20 general household member types - head (male) - wife - one child (male) - two children (male) - three children (male) - grandchild (male) - brother 16 Membership c ms C rare Composition - Single (male) - Single (female) - Couple - Head (female) + child (female) - Two other members - Couple + child (male) - Couple + child (female) - Couple + mother cms C rare c ms C rare - father - other (male) - two other members (male) - head (female) - one child (female) - two children (female) -three children (female) - grandchild (female) - sister - mother - child s wife - other (female) - two other members (female) - Three other members - Couple + two children (male) - Couple + child (male) + child (female) - Couple + two children (female) - four other members - Five members - Six members - Seven or more members : Correlation between Continuous Attributes : Rare Household Types 15

16 Parameter Estimation Parameters for Housing Type Choice Model Explanatory Variable Parameters (T-values) Own-detached Rent-detached Own-apartment Rent-apartment Number of household members 0.667(3.32) 0.645(2.72) 0.405(1.95) 0.503(2.47) Age of head 0.166(4.28) (-2.12) 0.106(2.64) 0.128(3.24) Sample size : 1,000 Log-likelihood : -1,118.6 Likelihood ratio : Parameters for Zone Choice Model Explanatory Variable Parameters (T-values) zone 1 zone 2-5 zone 6-8 Distance from the city center (-0.76) Number of household members (6.21) (7.23) Age of head (-2.45) Dummy of Own-detached (6.21) (2.39) Dummy of Own-apartment (-4.77) Dummy of Rent-apartment (5.14) (-1.78) Sample size : 1,000 Log-likelihood : Likelihood ratio :

17 Average Single (male) Single (female) Couple Head (female) + child (female) Two other members Couple + child (male) Couple + child (female) Couple + mother Three other members Couple + two children (male) Couple + child (male) + child (female) Couple + two children (female) four other members Five members Six members Seven or more members Standard deviation Estimation Results of Households by Member Composition Residuals from the Observed Data by Proposed and Naïve System (Own-detached) 300 Households 30 Households Average Proposed System (E1) Average Naive System (E2) Standard deviation Proposed System (E1) Standard deviation Naive System (E2) Observed Samples : , Proposed system appears to be slightly better than Naïve system 17

18 Observed Proposed System Naïve System Estimation Results of Households by Zone Households by Proposed and Naïve System (Couple + child (male) + child (female) ) zone6 zone5 zone2 zone zone1 zone7 0 0 zone4 zone8 - No significant improvement in proposed system - The result based on comparison with partial values 18

19 Goodness-of-Fit of the Estimation Proposed System Naïve System Average Standard Deviation Average Standard Deviation E E E E E E E E E E E E E E E E 平均 平均 Goodness-of-fit of Proposed system mostly better than Naïve system 19

20 Concluding Remarks Agent Based Household Data Estimation System for land-use microsimulation - Continuous and discrete attribute variables by agent including housing type and location - Validated with the case study Further Study - Development for other attributes (Car Ownership, income etc.) - Comparison with results by IPF method - Theoretical discussion for Monte Carlo Sampling - Application with urban model 20

21 Thank you 21