Atmospheric dynamics final project: A Trivial Test on a Bayesian Statistic Distribution Model for...

Atmospheric Dynamics Final Project: A Trivial Test on a Bayesian Statistic Distribution Model for Air PollutionWritten ReportB02602005顏東白This final project was a small test on the accuracy of a developing statistic distribution model. It was aimed to discuss further properties of the model for which may be used in future air pollution project.

1. TheoryThe model is a Bayesian model, assuming Gaussian distribution. The velocity on each observation site gives the expected position for the next time interval, while the variance is estimated from the data of all the observation sites.

Two key concepts were used to optimize the estimation: accuracy and creditability.

Accuracy refers to the minimization of the squared error:

∑j

(∑ φ ( x⃑ i , t n ) p ( x⃑ j− x⃑ i , v⃑ ( x⃑ i , t ))

∑ p ( x⃑ j− x⃑i , v⃑ ( x⃑ i , t ))−φ( x⃑ j , t n+1))

2

Creditability refers to maximization of the representative ability of the sample points to each other. It can be done by maximizing the logarithm of the likelihood function:

ln (L(σ2|(x i )n ) )=−n2 ln (2π σ 2)− 12σ2

∑ijd ij

2

Minimization of the square error yields

∑ijφ j d ij

2e−d ij

2

2σ 2 ∑ije

−dij2

2σ2 =∑ijφ j e

−dij2

2σ 2 ∑ijd ij

2e−dij

2

2σ 2

Maximization of the logarithm of the likelihood function yields:

σ̂ 2= 12n2

∑ijd ij2

An appropriate selection of sites is to select the sites for which

E [∑ijφ j d ij

2e−dij

2

2 σ̂2 ∑ije

−dij2

2 σ̂ 2 ]=E[∑ijφ j e

−dij2

2 σ̂ 2 ∑ijd ij

2e−dij

2

2 σ̂ 2 ]

Where sigma hat is the estimated standard deviation from MLE. Such selection was not used due to its complexity; instead a more naïve estimation was adopted. We tried to minimize the function

(− ln (2π σ2 )− 12n2σ2

∑ijdij2)2

+(∑ijφ j d ij

2 e−dij

2

2σ2 ∑ije

−dij2

2σ 2 −∑ijφ j e

−dij2

2σ2 ∑ijd ij2 e

−dij2

2σ 2 )2

As a tradeoff between MSE and MLE considerations.

2. Code

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

##Environment Setting

library("KernSmooth", lib.loc="C:/Program Files/R/R-3.2.2/library")

library("fields", lib.loc="~/R/win-library/3.2")

##Parameters Setting

#Gaussian Function

G<-function(x){

b=2.71828182846^(-x^2)/(2*pi)^0.5

return(b)

}

t=0.01

#random generator

x_0=rnorm(200,0,1)

y_0=rnorm(200,0,1)

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

31

data_0=cbind(x_0,y_0)

#process initial est.

dis_0=bkde2D(data_0,c(0.5,0.5),grid=c(1001L,1001L))

se=matrix(nrow=1001,ncol=1001,byrow=TRUE)

rse=matrix(nrow=1001,ncol=1001,byrow=TRUE)

for(i in 1:1001){

for(j in 1:1001){

se[i,j]=(G(dis_0$x1[i])*G(dis_0$x2[j])-dis_0$fhat[i,j])^2

rse[i,j]=((G(dis_0$x1[i])*G(dis_0$x2[j])-dis_0$fhat[i,j])/(G(dis_0$x1[i])*G(dis_0$x2[j])))^2

}

}

##Run advection using RK4

#velocity field

ode<-function(x){

32

33

34

35

36

37

38

39

40

41

42

43

44

45

46

47

m=matrix(ncol=2,nrow=1,byrow=TRUE)

m[1]=-.5*x[2]

m[2]=.5*x[1]

return(m)

}

#the Runge-Kutta Method

RK<-function(x){

k_1=ode(x)*t

k_2=ode(x+0.5*k_1)*t

k_3=ode(x+0.5*k_2)*t

k_4=ode(x+k_3)*t

x_f=x+(k_1+2*k_2+2*k_3+k_4)/6

return(x_f)

}

48

49

50

51

52

53

54

55

56

57

58

59

60

61

62

63

#Run Advection

data=matrix(ncol=2,nrow=200,byrow=TRUE)

data=data_0

for(i in 1:5000){

for(j in 1:200){

data[j,]=RK(data[j,])

}

}

#process final est.

dis=bkde2D(data,c(0.5,0.5),grid=c(1001L,1001L))

se=matrix(nrow=1001,ncol=1001,byrow=TRUE)

rse=matrix(nrow=1001,ncol=1001,byrow=TRUE)

for(i in 1:1001){

for(j in 1:1001){

se[i,j]=(G(dis$x1[i])*G(dis$x2[j])-dis_0$fhat[i,j])^2

64

65

66

67

68

69

70

71

72

73

74

75

76

77

78

79

rse[i,j]=((G(dis$x1[i])*G(dis_0$x2[j])-dis$fhat[i,j])/(G(dis$x1[i])*G(dis$x2[j])))^2

}

}

##Kriging Estimation

#Gaussian Function

G<-function(x){

b=2.71828182846^(-x^2)/(2*pi)^0.5

return(b)

}

t=.01

#input observed data

ns=31

obs=matrix(nrow=ns,ncol=3,byrow=TRUE)

80

81

82

83

84

85

86

87

88

89

90

91

92

93

94

95

for(i in 1:10){

x_obs[i]=cos(2*pi*i/10)

x_obs[i+10]=2*cos(2*pi*i/10)

x_obs[i+20]=.5*cos(2*pi*i/10)

y_obs[i]=sin(2*pi*i/10)

y_obs[i+10]=2*sin(2*pi*i/10)

y_obs[i+20]=.5*sin(2*pi*i/10)

}

x_obs[31]=0

y_obs[31]=0

d=matrix(nrow=length(x_obs),ncol=length(y_obs),byrow=TRUE)

sigma=c()

for(j in 1:length(x_obs)){

obs[j,1]=G(x_obs[j])*G(y_obs[j])

96

97

98

99

100

101

102

103

104

105

106

107

108

109

110

111

obs[j,2]=-.5*y_obs[j]

obs[j,3]=.5*x_obs[j]

}

for(j in 1:length(x_obs)){

for(k in 1:length(x_obs)){

d[j,k]=((x_obs[k]-x_obs[j])-obs[j,2]*t)^2+((y_obs[k]-y_obs[j])-obs[j,3]*t)^2

}

}

#MSE

f=c()

int=.001

L=410

U=440

c_1=matrix(nrow=length(x_obs),ncol=length(y_obs),byrow=TRUE)

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

c_2=matrix(nrow=length(x_obs),ncol=length(y_obs),byrow=TRUE)

c_3=matrix(nrow=length(x_obs),ncol=length(y_obs),byrow=TRUE)

c_4=matrix(nrow=length(x_obs),ncol=length(y_obs),byrow=TRUE)

for(i in L:U){

for(j in 1:length(x_obs)){

for(k in 1:length(y_obs)){

c_1[j,k]=obs[j,1]*d[j,k]*G(-sqrt(d[j,k]/(int*i)))

c_2[j,k]=G(-sqrt(d[j,k]/(int*i)))

c_3[j,k]=obs[j,1]*G(-sqrt(d[j,k]/(int*i)))

c_4[j,k]=d[j,k]*G(-sqrt(d[j,k]/(int*i)))

}

}

f[i]=(sum(c_1)*sum(c_2)-sum(c_4)*sum(c_3))^2

}

plot(seq(int*L,int*U,int),f[c(L:U)],main="Constrain Distribution",xlab="sigma",ylab="f(x)")

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

sigma=0.42 #sorting function requires

#MLE

sigma=sum(d)/2/ns^2

#hybrid

f=c()

g=c()

h=c()

int=.00001

L=43590

U=43600

c_1=matrix(nrow=length(x_obs),ncol=length(y_obs),byrow=TRUE)

c_2=matrix(nrow=length(x_obs),ncol=length(y_obs),byrow=TRUE)

c_3=matrix(nrow=length(x_obs),ncol=length(y_obs),byrow=TRUE)

c_4=matrix(nrow=length(x_obs),ncol=length(y_obs),byrow=TRUE)

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

for(i in L:U){

for(j in 1:length(x_obs)){

for(k in 1:length(y_obs)){

c_1[j,k]=obs[j,1]*d[j,k]*G(-sqrt(d[j,k]/(int*i)))

c_2[j,k]=G(-sqrt(d[j,k]/(int*i)))

c_3[j,k]=obs[j,1]*G(-sqrt(d[j,k]/(int*i)))

c_4[j,k]=d[j,k]*G(-sqrt(d[j,k]/(int*i)))

}

}

f[i]=(sum(c_1)*sum(c_2)-sum(c_4)*sum(c_3))^2

g[i]=(log(2*pi*(int*i)^2)+sum(d)/2/(int*i)/ns^2)^2

h[i]=g[i]+f[i]

}

plot(seq(int*L,int*U,int),h[c(L:U)],main="Constrain Distribution",xlab="sigma",ylab="h(x)")

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

sigma=.43595

#set grid

x_est=seq(-3,3,0.1)

y_est=seq(-3,3,0.1)

d_est=array(dim=c(length(x_est),length(y_est),length(x_obs)))

#estimation

est=matrix(nrow=length(x_est),ncol=length(y_est),byrow=TRUE)

err=matrix(nrow=length(x_est),ncol=length(y_est),byrow=TRUE)

rel=matrix(nrow=length(x_est),ncol=length(y_est),byrow=TRUE)

for(i in 1:length(x_est)){

for(j in 1:length(y_est)){

for(s in 1:length(x_obs)){

d_est[i,j,s]=((x_est[i]-x_obs[s])-obs[s,2]*t)^2+((y_est[j]-y_obs[s])-obs[s,3]*t)^2

}

176

177

178

179

180

est[i,j]=sum(obs[,1]*G(-sqrt(d_est[i,j,]/sigma)))/sum(G(-sqrt(d_est[i,j,]/sigma)))

err[i,j]=(G(x_est[i])*G(y_est[j])-est[i,j])^2

rel[i,j]=err[i,j]/(G(x_est[i])*G(y_est[j]))^2

}

}

3. Results and Discussion

1) The error in between the middle of sample points is the greatest.2) Relative error is acceptable only within the circle with radius 1, no matter how you select your sites.3) Site selection only alters the predicting ability within the r<1 circle.4) Inside r<0.5, prediction ability decreases; this can be improved when site selection is closer to origin.