1 | function ildp
|
---|
2 |
|
---|
3 | tic
|
---|
4 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
---|
5 | %pocatecni konstanty
|
---|
6 |
|
---|
7 | % [b=1, yr=1] - to jsme zkouseli spolu sigma = 0.1; rho = 0.5;
|
---|
8 | % [b=-1, yr=1] sigma = 0.1; rho = 0.5; !aprior
|
---|
9 | % [b=10, yr=5] sigma = 0.5; rho = 2.0;
|
---|
10 | % [b=0.6, yr=10] sigma = 0.1; rho = 1.0;
|
---|
11 |
|
---|
12 | Iterace = 10; %iterace
|
---|
13 | K = 5; %casy
|
---|
14 | N = 100; %vzorky
|
---|
15 |
|
---|
16 | sigma = 0.1;
|
---|
17 | Sigmas = [[sigma^2 0 0]; [0 0 0]; [0 0 0]];
|
---|
18 |
|
---|
19 | h = 0;
|
---|
20 | hdx = [0; 0; 0];
|
---|
21 | hdxdx = [[0 0 0]; [0 0 0]; [0 0 0]];
|
---|
22 |
|
---|
23 | Rk = 1*ones(1, K);
|
---|
24 | x0 = [0; 1; 1];
|
---|
25 |
|
---|
26 | %velikost okoli pro lokalni metodu
|
---|
27 | rho = 0.5;
|
---|
28 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
---|
29 | %globalni promenne
|
---|
30 |
|
---|
31 | Kpi = ones(4, K);
|
---|
32 | Kpi(4, :) = zeros(1, K);
|
---|
33 | % Kpi(3, :) = zeros(1, K);
|
---|
34 |
|
---|
35 | Wv = zeros(9, K);
|
---|
36 |
|
---|
37 | Xkn = zeros(3, K, N);
|
---|
38 | Xstripe = zeros(3, K);
|
---|
39 |
|
---|
40 | gka = 0;
|
---|
41 | gnu = 0;
|
---|
42 |
|
---|
43 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
---|
44 | %iteracni smycka
|
---|
45 |
|
---|
46 | for i = 1:Iterace,
|
---|
47 | %generovani stavu
|
---|
48 | for n = 1:N,
|
---|
49 | Xkn(:, 1, n) = x0 + [sigma*randn(); 0; 0];
|
---|
50 | for k = 1:K-1,
|
---|
51 | Uk = uPi(k, Xkn(:, k, n));
|
---|
52 | Kk = Uk*Xkn(3, k, n)/(Uk^2*Xkn(3, k, n) + sigma^2);
|
---|
53 | Xkn(1, k+1, n) = Xkn(1, k, n) + Xkn(2, k, n)*Uk + sigma*randn();
|
---|
54 | Xkn(2, k+1, n) = Xkn(2, k, n) + Kk*(Xkn(1, k+1, n) - Xkn(1, k, n) - Xkn(2, k, n)*Uk);
|
---|
55 | Xkn(3, k+1, n) = (1 - Kk*Uk)*Xkn(3, k, n);
|
---|
56 | end
|
---|
57 | end
|
---|
58 | Xstripe = mean(Xkn, 3);
|
---|
59 |
|
---|
60 | for k = K-1:-1:1,
|
---|
61 | gka = k;
|
---|
62 | % 1]
|
---|
63 | for n = 1:N,
|
---|
64 | %krive okoli
|
---|
65 | Xkn(1, k, n) = Xstripe(1, k) + rho*randn();
|
---|
66 | Xkn(2, k, n) = Xstripe(2, k) + rho*randn();
|
---|
67 | Xkn(3, k, n) = Xstripe(3, k)*exp(rho*randn());
|
---|
68 | end
|
---|
69 | % 2]
|
---|
70 | for n = 1:N,
|
---|
71 | gnu = n;
|
---|
72 | [Uopt(n), Hmin(n)] = fminunc(@Hamilt, uPi(k, Xkn(:, k, n)), optimset('GradObj','on','Display','notify'));
|
---|
73 | end
|
---|
74 | % 3]
|
---|
75 | for n = 1:N,
|
---|
76 | Vn(n) = Hmin(n) + Vtilde(k+1, Xkn(:, k, n));
|
---|
77 | end
|
---|
78 | % 4]
|
---|
79 | Epsilon = zeros(3, N);
|
---|
80 | for n = 1:N,
|
---|
81 | Epsilon(1, n) = Xkn(1, k, n) - Xstripe(1, k);
|
---|
82 | Epsilon(2, n) = Xkn(2, k, n) - Xstripe(2, k);
|
---|
83 | Epsilon(3, n) = Xkn(3, k, n)/Xstripe(3, k);
|
---|
84 | end
|
---|
85 | mFi = matrixFi(Epsilon);
|
---|
86 | FiFiTInvFi = (mFi*mFi')\mFi;
|
---|
87 | Wv(:,k) = FiFiTInvFi * Vn';
|
---|
88 | for n = 1:N,
|
---|
89 | yt(n) = Xkn(1, k, n);
|
---|
90 | bt(n) = Xkn(2, k, n);
|
---|
91 | pt(n) = Xkn(3, k, n);
|
---|
92 | end
|
---|
93 | mPsi = [yt',...
|
---|
94 | bt'.*Uopt',...
|
---|
95 | pt'.*Uopt',...
|
---|
96 | Uopt'];
|
---|
97 | PsiPsiTInvPsi = (mPsi'*mPsi)\mPsi';
|
---|
98 | Kpi(:,k) = PsiPsiTInvPsi * (Rk(k)*ones(N,1));
|
---|
99 | end
|
---|
100 | end
|
---|
101 | %%%%%%%%%%%
|
---|
102 | toc
|
---|
103 |
|
---|
104 | Kpi
|
---|
105 |
|
---|
106 | for n = 1:N,
|
---|
107 | Xkn(:, 1, n) = x0 + [sigma*randn(); 0; 0];
|
---|
108 | for k = 1:K-1,
|
---|
109 | Uk = uPi(k, Xkn(:, k, n));
|
---|
110 | UU(k,n) = Uk;
|
---|
111 | Kk = Uk*Xkn(3, k, n)/(Uk^2*Xkn(3, k, n) + sigma^2);
|
---|
112 | Xkn(1, k+1, n) = Xkn(1, k, n) + Xkn(2, k, n)*Uk + sigma*randn();
|
---|
113 | Xkn(2, k+1, n) = Xkn(2, k, n) + Kk*(Xkn(1, k+1, n) - Xkn(1, k, n) - Xkn(2, k, n)*Uk);
|
---|
114 | Xkn(3, k+1, n) = (1 - Kk*Uk)*Xkn(3, k, n);
|
---|
115 | end
|
---|
116 | hold all
|
---|
117 | subplot(4,1,1);
|
---|
118 | plot(1:K,Xkn(1,:,n))
|
---|
119 | hold all
|
---|
120 | subplot(4,1,2);
|
---|
121 | plot(1:K,Xkn(2,:,n))
|
---|
122 | hold all
|
---|
123 | subplot(4,1,3);
|
---|
124 | plot(1:K,Xkn(3,:,n))
|
---|
125 | hold all
|
---|
126 | subplot(4,1,4);
|
---|
127 | plot(1:K-1,UU(:,n))
|
---|
128 |
|
---|
129 | end
|
---|
130 | title('iLDP')
|
---|
131 | figure
|
---|
132 | for n = 1:N,
|
---|
133 | Xkn(:, 1, n) = x0 + [sigma*randn(); 0; 0];
|
---|
134 | for k = 1:K-1,
|
---|
135 | Uk = (Rk(k) - Xkn(1, k, n))/(Xkn(2, k, n) + Xkn(3, k, n));
|
---|
136 | UU(k,n) = Uk;
|
---|
137 | Kk = Uk*Xkn(3, k, n)/(Uk^2*Xkn(3, k, n) + sigma^2);
|
---|
138 | Xkn(1, k+1, n) = Xkn(1, k, n) + Xkn(2, k, n)*Uk + sigma*randn();
|
---|
139 | Xkn(2, k+1, n) = Xkn(2, k, n) + Kk*(Xkn(1, k+1, n) - Xkn(1, k, n) - Xkn(2, k, n)*Uk);
|
---|
140 | Xkn(3, k+1, n) = (1 - Kk*Uk)*Xkn(3, k, n);
|
---|
141 | end
|
---|
142 | hold all
|
---|
143 | subplot(4,1,1);
|
---|
144 | plot(1:K,Xkn(1,:,n))
|
---|
145 | hold all
|
---|
146 | subplot(4,1,2);
|
---|
147 | plot(1:K,Xkn(2,:,n))
|
---|
148 | hold all
|
---|
149 | subplot(4,1,3);
|
---|
150 | plot(1:K,Xkn(3,:,n))
|
---|
151 | hold all
|
---|
152 | subplot(4,1,4);
|
---|
153 | plot(1:K-1,UU(:,n))
|
---|
154 |
|
---|
155 | end
|
---|
156 | title('CE')
|
---|
157 |
|
---|
158 | for vzorek = 1:100,
|
---|
159 | loss(vzorek) = 0;
|
---|
160 | bb = randn() + x0(2);
|
---|
161 | yy(1) = x0(1);
|
---|
162 | for k=1:K-1,
|
---|
163 | yy(k+1)=yy(k)+bb*uPi(k,[yy(k); bb; 0]);
|
---|
164 | loss(vzorek) = loss(vzorek) + (yy(k+1) - Rk(k+1))^2;
|
---|
165 | end
|
---|
166 | end
|
---|
167 | figure
|
---|
168 | hist(log(loss))
|
---|
169 |
|
---|
170 | %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
---|
171 | %pomocne funkce
|
---|
172 |
|
---|
173 | function [val_uPi] = uPi(k_uPi, x_uPi)
|
---|
174 | val_uPi = (Rk(k_uPi) - Kpi(1, k_uPi)*x_uPi(1))/(Kpi(2, k_uPi)*x_uPi(2) + Kpi(3, k_uPi)*x_uPi(3) + Kpi(4, k_uPi));
|
---|
175 |
|
---|
176 | end
|
---|
177 |
|
---|
178 | function [val_ham, val_grad] = Hamilt(u_ham)
|
---|
179 | % Vtddx = Vtilde_dx(gka+1, Xkn(:, gka, gnu));
|
---|
180 | val_ham = (Xkn(1, gka, gnu) + Xkn(2, gka, gnu)*u_ham - Rk(gka+1))^2 + Xkn(3, gka, gnu)*u_ham^2 ... + sigma^2 ... %ztrata l
|
---|
181 | + [Xkn(2, gka, gnu)*u_ham; ...
|
---|
182 | Xkn(3, gka, gnu)*u_ham*(Xkn(1, gka+1, gnu) - Xkn(1, gka, gnu) - Xkn(2, gka, gnu)*u_ham)/(Xkn(3, gka, gnu)*u_ham^2 + sigma^2); ...
|
---|
183 | -Xkn(3, gka, gnu)^2*u_ham^2/(Xkn(3, gka, gnu)*u_ham^2 + sigma^2)]' ... %fce f
|
---|
184 | *Vtilde_dx(gka+1, Xkn(:, gka, gnu)) ...
|
---|
185 | + Wv(5, gka+1)*sigma;%+ Wv(4, gka+1)*sigma;
|
---|
186 |
|
---|
187 | val_grad = 2*(Xkn(1, gka, gnu) + Xkn(2, gka, gnu)*u_ham - Rk(gka+1))*Xkn(2, gka, gnu) + 2*Xkn(3, gka, gnu)*u_ham ... %ztrata du
|
---|
188 | + [Xkn(2, gka, gnu);...
|
---|
189 | (2*u_ham^2*Xkn(3, gka, gnu)^2*(Xkn(1, gka, gnu) - Xkn(1, gka+1, gnu) + u_ham*Xkn(2, gka, gnu)))/(sigma^2 + u_ham^2*Xkn(3, gka, gnu))^2 - (u_ham*Xkn(2, gka, gnu)*Xkn(3, gka, gnu))/(sigma^2 + u_ham^2*Xkn(3, gka, gnu)) - (Xkn(3, gka, gnu)*(Xkn(1, gka, gnu) - Xkn(1, gka+1, gnu) + u_ham*Xkn(2, gka, gnu)))/(sigma^2 + u_ham^2*Xkn(3, gka, gnu));...
|
---|
190 | (2*u_ham^3*Xkn(3, gka, gnu)^3)/(sigma^2 + u_ham^2*Xkn(3, gka, gnu))^2 - (2*u_ham*Xkn(3, gka, gnu)^2)/(sigma^2 + u_ham^2*Xkn(3, gka, gnu))]' ... %fce f du
|
---|
191 | * Vtilde_dx(gka+1, Xkn(:, gka, gnu)); %derivace Bellman fce
|
---|
192 | end
|
---|
193 |
|
---|
194 | function [val_Vt] = Vtilde(k_Vt, x_Vt)
|
---|
195 | if(k_Vt == K)
|
---|
196 | val_Vt = h;
|
---|
197 | else
|
---|
198 | Epsl = zeros(3, 1);
|
---|
199 | Epsl(1) = x_Vt(1) - Xstripe(1, k_Vt);
|
---|
200 | Epsl(2) = x_Vt(2) - Xstripe(2, k_Vt);
|
---|
201 | Epsl(3) = x_Vt(3)/Xstripe(3, k_Vt);
|
---|
202 |
|
---|
203 | val_Vt = vectFi(Epsl)' * Wv(:,k_Vt);
|
---|
204 | end
|
---|
205 | end
|
---|
206 |
|
---|
207 | function [val_Vt] = Vtilde_dx(k_Vt, x_Vt)
|
---|
208 | if(k_Vt == K)
|
---|
209 | val_Vt = hdx;
|
---|
210 | else
|
---|
211 | Epsl = zeros(3, 1);
|
---|
212 | Epsl(1) = x_Vt(1) - Xstripe(1, k_Vt);
|
---|
213 | Epsl(2) = x_Vt(2) - Xstripe(2, k_Vt);
|
---|
214 | Epsl(3) = x_Vt(3)/Xstripe(3, k_Vt);
|
---|
215 |
|
---|
216 | val_Vt = difFi(Epsl)' * Wv(:,k_Vt);
|
---|
217 | end
|
---|
218 | end
|
---|
219 |
|
---|
220 | function [val_Fi] = vectFi(x_Fi)
|
---|
221 | val_Fi = [ ...
|
---|
222 | 1; ...
|
---|
223 | x_Fi(1); ...
|
---|
224 | x_Fi(2); ...
|
---|
225 | log(x_Fi(3)); ...
|
---|
226 | x_Fi(1)^2; ...
|
---|
227 | x_Fi(1)*x_Fi(2); ...
|
---|
228 | x_Fi(1)*log(x_Fi(3)); ...
|
---|
229 | x_Fi(2)^2; ...
|
---|
230 | x_Fi(2)*log(x_Fi(3)); ...
|
---|
231 | % 2*ln(x_Fi(3)); ...
|
---|
232 | ];
|
---|
233 | end
|
---|
234 |
|
---|
235 | function [val_Fi] = matrixFi(x_Fi)
|
---|
236 | val_Fi = [ ...
|
---|
237 | ones(1, N); ...
|
---|
238 | x_Fi(1, :); ...
|
---|
239 | x_Fi(2, :); ...
|
---|
240 | log(x_Fi(3, :)); ...
|
---|
241 | x_Fi(1, :).^2; ...
|
---|
242 | x_Fi(1, :).*x_Fi(2, :); ...
|
---|
243 | x_Fi(1, :).*log(x_Fi(3, :)); ...
|
---|
244 | x_Fi(2, :).^2; ...
|
---|
245 | x_Fi(2, :).*log(x_Fi(3, :)); ...
|
---|
246 | ];
|
---|
247 |
|
---|
248 | end
|
---|
249 |
|
---|
250 | function [val_Fi] = difFi(x_Fi)
|
---|
251 | val_Fi = [ ...
|
---|
252 | 0 0 0; ...
|
---|
253 | 1 0 0; ...
|
---|
254 | 0 1 0; ...
|
---|
255 | 0 0 1/(x_Fi(3)); ...
|
---|
256 | 2*x_Fi(1) 0 0; ...
|
---|
257 | x_Fi(2) x_Fi(1) 0; ...
|
---|
258 | log(x_Fi(3)) 0 x_Fi(1)/(x_Fi(3)); ...
|
---|
259 | 0 2*x_Fi(2) 0; ...
|
---|
260 | 0 log(x_Fi(3)) x_Fi(2)/(x_Fi(3)); ...
|
---|
261 | ];
|
---|
262 | end
|
---|
263 | end |
---|