以「类神经网路」结合「细胞自动机」概念之地理资讯系统来分析台重点

1、Artificial Neural Network and Cellular Automata As A Modeli ng Simulation for Night Market Spatial Developme ntShwu-Ti ng Lee, Tsu-Chia ng Lei, Chih-We n WuFen g-Chia Uni versity, Departme nt of Architecture, .twAbstract : Night-market is a special liv ing culture of Taiwa n. It is a n a

2、tural process of spatial developme nt, and prese nts popular ni ght en terta inment of Taiwa nese people. The goal of this research is to establish the simulation model to analyze spatial developme nt factors of Taiwa nese Night-market. The method is in tegratedCellular Automata(CA andArtificial Neu

3、ral Network(ANN to experiment spatial simulatio n of Night-market. The simulatio n process is employed by Geographic Information System (GIS. The result of this research illustrates the 76.897% similarity betwee n simulatio n result and existi ng con diti on. It provides the evide nee that the imple

4、me nt simulatio n parameters have the appare nt conn ecti ons betwee n Taiwa nese Night-market spatial developme nt and street structures.Key words: Geographic In formation System (GIS, Artificial Neural Network (ANN, Cellular Automata (CA, Taiwa nese ni ght market.1. In troducti onNight-market is a

5、 unique culture in Taiwan. It combines traditional culture and con sumptio n behavior. The followi ng are some clarificati ons of Taiwa nese ni ght-market:(1. Night-markets are kinds of street-market which operating mostly at night. (Tsai, 1995(2. Night-markets are consisted of stores and stalls alo

6、ng the street. The activity at ni ght is Ion ger tha n daytime. (Hou, 1988(3. Night-markets are the aggregati on places of stalls, and people. It sells all kinds of goods and food. There are some features of ni ght-market in cluded casual, convenient, bustli ng, and family-suited. (Yu, 1992(4. Night

7、-markets are the aggregati on place of stores and stalls which are regularly operating at night with large scale. (Wang, 1996Many stalls of n ight-market don ' t have legal bus in ess lice nses and they run the bus in ess from local union agreeme nt. Taiwa nese gover nment had regarded ni ght-ma

8、rkets as the urba n chaotic problem formerly. Yu men ti oned n ight-markets reveal characters of undeveloped ' situation under the ideology of modernization in Taiwan. The chaos of ni ght-market con tradicts the aim of being n eat and in order emphasized by moder n desig n (Yu, 1994. Data from T

9、ourism Bureau at 2004 shows that Taiwa nese ni ght-market has drow n great atte nti on for foreig n tourists. Therefore, the gover nment has listed ni ght market as the importa nt developme nt project in the Tourists D,c20l04,. 2008' (Srece nt years for solvi ng the chaotic problems of ni ght-ma

10、rket, gover nment restricted many special-purpose zones for the ni ght markets. Stalls could operate the bus in ess legally in thesezones. However, after being changed by modernized value as sightseeing ” areas, many ni ght markets had lost their orig inal characters. Being crowded is probably the r

11、eas on that ni ght market is regarded as the most popular ni ght en terta inment at public open space (Yu, 1994. Consequently, the policy of setting special-purpose zone for night market usually causes the decli ne of ni ght-market developme nt, and los ing orig inal characters and attract ions of n

12、 ight-market. Therefore the developme nt of ni ght market is stro ngly conn ected with the local street structure and its con text.2. Research methodThis research establishes the simulati on model of ni ght-market spatial developme nt by applyi ng methodology of CA and ANN on GIS platform.2.1 The co

13、ncept of Cellular Automata (CAThe con cept of CA was occurred arou nd 19201930 (Almeida and Gleria ni, 2005.Evoluti on of CA shows self-duplicati on in the form of unit cells. It also shows selfsimilarity and self-orga ni zati on. (Wolfram, 1994.CA model is contained four comp onents as followed: ce

14、lls, states, n eighborhood, and rule. (Batty et al., 1997.1. Cells: the form of cellsCA is con stituted by group of grids (or cells. These grids may be any geometric shapes, and also can be three dime nsional objects. However, the regular square grid is most exte nsively used in CA releva nt researc

15、h. (Fig. 1Fig. 1 The cells setting of CA:Left figure expresses check grid state, and right figure expresses hexag on grid state.2. States: evolved state of cellsThe state of each cell can be bin ary, for example alive or dead, vacuum or occupied.It could also be multiple, for example residential lan

16、d-use commercial land-useleisure Iand-use factory Iand-useand etc. (Fig. 2Fig. 2 The states setting of CA:0: white, dead, and un developed; 1: black, alive, and developed.3. Neighborhood: relati on with a sin gle cell's n eighborsThe state of each cell evolves according to the neighbors' sta

17、tes. The region of the neighbor size needs to be defined in a CA model to determine the scale of interaction. (Fig. 3Fig. 3 Neighborhood setting of CA:B and A have the adjacency, but C and A don' t have. Assumption A is thedeveloped cell, and the n follow ingdevelopment situation is that B cell

18、has more developed opportunity than C cell.4. Rule: evoluti on rule of cellsThe state of each cell in the next step is determined by the interaction with neighbor region and their present states. CA's evolution is determined by certain defined rules. (Fig.4Fig. 4 Rule setting of CA:Diagram A exp

19、la ins the con diti on of con sta nt evolved process. Diagram B expla ins the con diti on of cell dying of Ion ely or crowded in evolved process. Diagram C expla ins the con diti on of cell breedi ng in evolved process.CA model is established accord ing to the upper four basic sett in gs. With GIS t

20、ech no logy progressive, many applicati ons in urba n an alyz ing used method of in tegrati on of CA and GIS (Couclelis, 1994, 1997; Batty and Xie, 1994; White and En gele n, 1993; Wu and Webster, 1998.This research used CA evolution model which is based on GIS platform to simulate and an alyze ni g

21、ht-market spatial developme nt. The setti ngs of CA- Rule and CA- Neighborhood are decided in Artificial Neural Network (ANN of the followi ng in troductio n.2.2 The con cept of Artificial Neural Network (ANNArtificial Neural Network (ANN is a very useful mathematical tool for percepti on and classi

22、ficatio n. It is through training and lear ning process to decide the classificati on of new samples. Fig. 5 is the diagram to explain concept of ANN implementing. ANN is con sisted of three parts: in put layer, hidde n layer, and output layer. Any ANN must experience the training step, then to prod

23、uce classified results.S1S5: Sample (training-samples or new-samples parameters T: Training target A:Output result W: WeightsInput LayerHidden Layeroutput LayerIsi :52 :53 :IInput LayerHidden Layerb: Bias Fig. 5 Diagram of ANN con cept:output Layer2.3 Integration of CA and ANNWorks associated ANN to

24、 CA models for spatial an alysis are quite few. Li and Yeh (2001 con ducted a simulati on of Ian d-use cha nged for a cluster of cities in southern China. ANN embedded in a CA model upon a binary state basis (urba n/nonurban use was tested in that experime nt. They further refi ned this model deali

25、ng with multiple regi on al la nd- use (Yeh and Li, 2002, and simulati on for alter native developme nt sce narios (Yeh and Li, 2003.This research is an application of ANN to decide the evolved rules of CA-model in simulatio n process of ni ght-market. The inv estigated area is tran sformed in to CA

26、 grids based on GIS platform. It is divided into training-area and simulation-area. The samples in training-area are used to train ANN, and the trained ANN is used to analyze samples in simulation-area in order to determine the evolution probabilities. These evolution probabilities structured the ru

27、les of CA model in this research.Input LayerHidden Layersi3. Research area and materials3.1 Research areaThe simulation target of this research is Feng-Chia night-market in Taichung, Taiwan. Night-market is usually occurred from a public aggregation site such as schools and large temples. Examples l

28、ike Shi-L in ni ght-market in Taipei arose from Tsz-Shia n temple. Then after Mine-Chuan University relocated to this area in 1963, it stimulated the expansion of Shi-Lin night market (Yu, 1994. The other example Feng-Chia night-market thrived because of the n eighborhood around Fen g-Chia Uni versi

29、ty. It becomes prosperous after years of developme nt and is one of the biggest ni ght-market in Taiwa n. The following Fig. 6 defines the investigation domain:A. Feng-Chia University 1. No. 12-99, Wenhua Road Training area 2. No. 5-12, Wenhua Road Trai ning area 3. Fe ng-Chia Roa Trai ning area 4.

30、No. 101-129, Wen hua Road Simulati on area 5. Lane 20, Fen g-Chia RoaSimulati on area 6. Feng- Chia Road Simulatio n area 7. Fen g-Chia Roa Simulati on area Fig. 6 Expla nati on on research doma inFen g-Chia ni ght-market is a typical n ight-market in Taiwa n. The stalls aggregate along the streets

31、in the night. In the earlier period, the night market only covered the area in dicated in secti on-1 (Fig. 6 . It gradually expa nded to no wadays exte nt. This study model divided the streets of Feng-Chia nightmarket as trainairea ' and simulationarea ' for the purpose of training process i

32、n ANN and simulation process in CA after trai ning.3.2 CA model set-upThe CA model sett ings are grid-sett ing, state-sett ing, and rule-sett ing:A. Grid-setti ng: Like many other ni ght markets, the commercial behavior of Feng- Chia night-market developed along the street. There are 2 types of stal

33、ls: one is immobile and the other is mobile. The immobile stalls are those stores in side the build ings on both sides of the street in dicate in C sect ion (Fig. 7. The mobile stalls are those occupy places in front of build ing arcade or sidewalk on both sides of the street. It is in dicated in se

34、ct ionB (Fig. 7. Mobile stalls usually rent the place from immobile stalls to obtain the trading permissi on. Thispermissi on does n ' t have any legitimate and is just a com mon consen sus of the local stalls and local union.Fig. 7BIThe cell grid-setting for streetThe expla natio n of CA grid-s

35、ett ing above is:1. Each cell is a grid form of 2m by 2m. It is the smallest size of operating stall in Fen g-Chia ni ght-market.2. Grid-1 and grid-9 are the spaces in side the buildi ngs. Grid-2 and grid-8 are the arcade or sidewalk spaces in front of the buildi ngs. Grid-3 to grid-7 are the spaces

36、 of street.3. The target of this research is to observe the development state on the streets of night-market. Therefore these grids are the simplified states from existing condition.B. State-setti ng: There are two states in CA model evoluti on: developed and un developed (Fig. 8.The explanation of

37、CA state-setting above is: 1. Grid-1 is the developed state, andit ' s the immobile stalls. 2. G2ds the developed state, and it' s the mobile stalls. 3.Grid-3 is the undeveloped state, and it' s the mobile-4t0dlth guGdid/elopedstate and it ' s the grid of road.Fig. 8 The cell state-s

38、ett ing for streetC. Rule-setting: The procedure of this research applied ANN to decide thedevelopment probability for each grid in the process of CA evolution. The results of CA evolutio n are dem on strated on GIS. The process exploited the trainin g-area in cludi ngsection-1, section-2, and secti

39、on-3 as ANN training-sample and training-target. Thesimulati on-area is sectio n-4,sect ion-5, sectio n-6, and sect ion-7. It was set in the trained ANN to gen erate the development probability for each grid. These probabilities are the setting of CA-rule.3.3 Establish parameters for ANNThe data of

40、each grid is tran sferred to the followi ng parameters for ANN operati on.The values of parameters are through the continual process of try-a nd-error. The process is through the comparis on and adjustme nt of prese nt data and simulatio n data. This result represe nts parameters' sig nifica nee

41、 and corresp onding relatio nships. Table 1: Parameter and definitionParameter D Rw Rs Ri SDefinition the index of core the index of road the index of road the index of road the in dex of stalldepth width positi on in tersecti on gatheri ngD : The in dex of core depth. The core located on sect ion-1

42、 is the beg inning place of Fen g-Chia n ight-market. The D value decreases as grid farther away from the core. We found in the process of simulation that decreasing in D value follows the pattern of streets in stead of radial dista nce from the core.9-a experimented the D values based on the settin

43、g of radial decreasing from the core. 9-b experimented the D values based on the setting of distance decreasing along the street structure from the core. The research process discovered that the setti ng of D values by 9-b is more appropriate through the comparison of existing condition.Fig. 9 Param

44、eter: D setti ngTable 2: Districts of parameter DTable 3: Values of parameter DDistrict Location D1 value D2 value D3 value D1 The streets of section-1 andsection-3. D1-a 7 D2-a 5.5 D3-a 3.5 D2 The streets of section-4 and section-5. D1-b 6 D2-b 3.5 D3-b 2 D3 The streets of sectio n-2, sectio n-6, a

45、nd sectio n-7. D1-c 5 D2-c 2.5 D3-c 0.5 D1 is the primary core area. D320M width road. Its development is depended on the aggregatio n effect of stalls rather tha n the affecti on by core area. D2 is a sec on dary area that adjace nt to D1. Its developme nt is stron gly affected by D1. The joints of

46、 D1 and D2 are the road in tersecti ons. The commercial activities at the in tersect ion show the tendency of in crease.Rw : The in dex of road width. This in dex in dicates the road width that has affected the aggregati on of stalls in ni ght-market: The aggregatio n of stalls is lower in wider roa

47、d and the higher in n arrower road.Table 4: Values of parameter RwRoad width 0m v Rw v 5m 5m v Rw v 10m 10m v Rw v 15m 10m v Rw v 20m Rw Value Rw = 2.5 Rw = 2 Rw = 1 Rw = 0Rs : The in dex of road positi on. This in dex in dicates the locati ons on street. These locations may have positive or negativ

48、e development factors. Grid-1 and grid-9 are the locations of immobile stalls with legitimacy. Therefore, these grids show the tendency of positive developme nt. Although grid-2 and grid-8 may have higher developme nt with mobile stalls, but these locations need to pay rental and still struggle with

49、 legality. Grid-3 to grid-7 are the locati ons of road, and have strong tendency of n egative developme nt.Fig. 10 Parameter: Rs setti ng Table 5: Values of parameter RsRs = 1 Rs = -1 Rs = -5Table 6: V alues of parameter RiLocatio n at road in tersecti on Yes NoRi Value Ri = 1 Ri = 0Ri : The in dex

50、of road in tersecti on. The positive developme nt tendency of stalls in road in tersect ion. S : The in dex of stalls gatheri ng. The grid adjace nt to the other stalls has higher developme nt probability.Fig. 11 The differe nt types of S -valuesWith no adjace nt stall grid-A in dicates S =0. With o

51、ne adjace nt stall grid-A in dicates that S =1. With two adjace nt stalls grid-A in dicates that S =2. With three adjace nt stalls grid-A in dicates that S =3. With four adjace nt stalls grid-A in dicates that S =4.4. Result of simulation experimentThis research used Arcmap 9.0 to execute the CA sim

52、ulatio n process. It allows obtaining the location attributes of GIS and analyze the process of night market simulatio n experime nt. I n ANN process, Matlab2008 is applied to con duct trai nin g-stage and gen erated developme nt probability in simulatio n stage.Fig. 12 Research structure4.1 Process

53、 of simulation experimentA. Curre nt developme nt data tran sfer CA grid-sett ing in trainin g-area secti on-1, secti on-2, and secti on-3. All grids with parameters are in put to ANN trainin g-stage. In thefollow ing, Fig. 13 displays the fini shed CA grid-13-3.setti ng in trainin g-area.Fig. 13 CA

54、 grid-sett ing in trainin g-area:13-1 is the trainin g-area sect ion-1. 13-2 is the trainin g-area secti on-2. 13-3 is the trainin g-area secti on-3.B. Probability is gen erated by fini shed trai nin g-ANN through the in putted parameters of grids in simulati on-area sectio n-4, sect ion-5, sectio n

55、-6, and sect ion-7. In the reality, there will not have the same development probability at two different locati ons with same con diti ons. Therefore the additi on ally in put of the ran dom probability is betwee n -2.5% +2.5% owing to the final development probability.Fig. 14 Current data of secti

56、on-4 and section-5 in simulation-area.C. Simulation result of stage one: There are 1,173 grid-data in simulation-area. The matched grids nu mber is 902from the comparison of simulation result and current data. The similarity average is76.897%.Fig. 15 Simulation result of stage one: 15-a is the simul

57、ation- area 4 and 5, 15-b is the simulati on area 6 and 7.Simulation result of stage two: With the result from stage-one simulation continues to con duct sec ond simulati on experime nt. The result from sec ond simulati on shows high capacity developme nt. The primary factor for spatial developme nt

58、 on sec ond simulati on result is determ ined by parameter S.Fig. 16 Simulation result of stage two: 16-a is the simulation-area 4 and 5, 16-b is the simulati on-area 6 and 7. E. Simulati on result of stage three: The purpose of third simulation experiment is to mend the situation of over-developed stalls on street. While rental will get higher with increasingly commercial activities and the legitimacyissues, the parameter Rs value of mobile stalls is mani pula