1 | \section{Nutnost p� k suboptim��metod� |
---|
2 | A�liv dynamick� programov� p����ok v ��lohy stochastick� ��analytick�e�en�bvykle nen�o�n��at. V ka�d��ov�kroku se toti� pot�se dv� obecn�bt��obl�my: 1) v� st� hodnoty vzhledem k $w_k$ a 2) n�edn�inimalizace vzhledem k $u_k$. Oba probl� obecn�emaj�nalytick�e�en� bez dal��pecifikace � je proto t�p� k aproxima�m metod� |
---|
3 | |
---|
4 | \section{Du���n� |
---|
5 | �stou situac� � stochastick� ��e, �e syst�popsan��m rovnic \eqref{sys} obvykle z�s�a n�k�parametru $\theta$, o kter�m� k dispozici pouze n�kou apriorn�nformaci. K �n� ��e tedy vhodn�ejen inimalizovat aktu��tr�, ale rovn�z�at o syst� co nejv� informac�ro minimalizaci budouc� ztr� Tento postup se naz����n�ref]. |
---|
6 | \subsection{Bayesovsk��� |
---|
7 | P�ar�up jak pro parametr $\theta$ z�at aposteriorn�ustotu pravd�dobnosti $f(\theta|X)$, je-li k dispozici apriorn�ustota pravd�dobnosti $f(\theta)$ a soubor m�n�X$, je aplikace Bayesova vzorce |
---|
8 | \begin{equation} |
---|
9 | \label{bay} |
---|
10 | f(\theta|X)=\frac{f(X|\theta)f(\theta)}{\int f(X|\theta)f(\theta)\mathrm{d}\theta} |
---|
11 | \end{equation} |
---|
12 | Rekurzivn�ou�it�zorce \eqref{bay} pro odhad parametru $\theta$ je postup Bayesovsk� u��ref]. |
---|
13 | |
---|
14 | P�nkr��vypo� m��ak tento p�p dv�ev� 1) nikdy nem� k dispozici $f(X|\theta)$, ale pouze jej�proximaci z m�n� a 2) aposteriorn�ustota pravd�dobnosti nemus��analytick�yj�en�co� jej�ou�it� dal��v� komplikuje. |
---|
15 | |
---|
16 | \subsection{Kalman�ltr} |
---|
17 | Pokud je p�tem ��yst� s gausovk�em, ve kter�nezn� parametry vystupuj�ako line� �ny situace se zna� zjednodu��ref]. Syst�\eqref{sys} m� �e $t$ tedy tvar |
---|
18 | |
---|
19 | \begin{equation} |
---|
20 | \label{sys2} |
---|
21 | x_{t+1}=f_k(x_t,u_t)+A_t(x_t,u_t)\theta_t+w_t |
---|
22 | \end{equation} |
---|
23 | |
---|
24 | , kde $A_t(x_t,u_t)$ je zn� matice z�s� na p�oz�stavu syst� a vstupu. D� p�kl�jme gausovsk�ozlo�en�umu $w_t$ se zn�m rozptylem, gausovsk�ozlo�en�ezn�ho parametru $\theta$ a jejich nekorelovanost, tedy |
---|
25 | \begin{equation} |
---|
26 | \theta_t\sim N(\hat{\theta},P_t), |
---|
27 | \end{equation} |
---|
28 | \begin{equation} |
---|
29 | w_t\sim N(0,Q_t), |
---|
30 | \end{equation} |
---|
31 | \begin{equation} |
---|
32 | \cov(w_t,\theta_t)=0. |
---|
33 | \end{equation} |
---|
34 | |
---|
35 | Na z�ad�dezvy syst� $x_{t+1}$ a $\theta_t$ chceme z�at n�k� odhad parametru $\theta_{t+1}$. Budeme p�kl�t, �e $\theta_{t+1}$ z�� line��pravou $\theta_t$ �ou neur�osti v syst�. Tedy �e |
---|
36 | \begin{equation} |
---|
37 | \label{opr} |
---|
38 | \hat{\theta}_{t+1}=\hat{\theta}_t+K_t(x_{t+1}-f_t(x_t,u_t)-A_t\hat{\theta}_t) |
---|
39 | \end{equation} |
---|
40 | , kde $K_t$ je nezn� matice, kterou ur�e z po�adavku minimalizace v��atice rozptylu $P_{t+1}$. Pro ni jako funkci $K_t$ m� ps� |
---|
41 | \begin{equation} |
---|
42 | P_{t+1}(K_t)=\E[(\theta-\hat{\theta}_{t+1})(\theta-\hat{\theta}_{t+1})^T]. |
---|
43 | \end{equation} |
---|
44 | Dosazen�za $\hat{\theta}_{t+1}$ ze \eqref{opr} a za $x_t$ ze \eqref{sys2} a �ou dostaneme (pro libovolnou matici $B$ budeme pro lep��itelnost nam�o $BB^T$ ps�zkr�n�B^2$) |
---|
45 | \begin{align} |
---|
46 | P_{t+1}(K_t)&=\E[(\theta-\theta_t-K_t(x_{t+1}-f_t(x_t,u_t)-A_t\hat{\theta}_t))^2] \nonumber \\ |
---|
47 | &=\E[((I-K_tA_t)(\theta-\theta_t)-K_tw_t)^2] \nonumber \\ |
---|
48 | &=(I-K_tA_t)\E[(\theta-\theta_t)^2](I-K_tA_t)^T-(I-K_tA_t)\cov(\theta,w_t)K_t^T-\nonumber \\ |
---|
49 | &-K_t\cov(\theta,w_t)(I-K_tA_t)^T+K_t\E[w_t^2]K_t^T. |
---|
50 | \end{align} |
---|
51 | |
---|
52 | Pou�it�definice $P_t$, $Q_t$ a p�kladu $\cov(\theta,w_t)=0$ m� |
---|
53 | \begin{equation} |
---|
54 | \label{Pt+1} |
---|
55 | P_{t+1}(K_t)=(I-K_tA_t)P_t(I-K_tA_t)^T+K_tQ_tK_t^T. |
---|
56 | \end{equation} |
---|
57 | Proto�e po�adujeme minim��ozptyl odhadu $\hat{\theta}_{t+1}$, ur�e $K_t$ z rovnice |
---|
58 | \begin{equation} |
---|
59 | \frac{\partial \tr( P_t)}{\partial K_t}. |
---|
60 | \end{equation} |
---|
61 | K proveden�derivace pou�ijeme vzorce*ODVOZENI BUDE ASI AZ V DODATKU* |
---|
62 | \begin{equation} |
---|
63 | \frac{\partial\tr(MXN)}{\partial X}=M^TN^T |
---|
64 | \end{equation} |
---|
65 | \begin{equation} |
---|
66 | \frac{\partial\tr(MXNX^TO)}{\partial X}=M^TO^TXN+OMXN. |
---|
67 | \end{equation} |
---|
68 | T�z�� line��ovnici pro $K_t$ tvaru |
---|
69 | \begin{equation} |
---|
70 | -P_t^TA_t-P_tA_t+K_tA_tP_tK_t+K_tA_t^TP_tK_t+2QK_t=0, |
---|
71 | \end{equation} |
---|
72 | kter��e�en�\begin{equation} |
---|
73 | \label{Kt} |
---|
74 | K_t=\frac{P_tA_t}{A_t^TP_tA_t+Q} |
---|
75 | \end{equation} |
---|
76 | Dosazen�\eqref{Kt} do \eqref{Pt+1} po uprav�ostaneme |
---|
77 | \begin{equation} |
---|
78 | \label{Pt+12} |
---|
79 | P_{t+1}=(I-K_tA_t)P_t |
---|
80 | \end{equation} |
---|
81 | Celkov�edy od p�� odhadu parametru $N(\hat{\theta}_t,P_t)$ k nov� $N(\hat{\theta}_{t+1},P_{t+1})$ p�me pomoc�\begin{equation} |
---|
82 | K_t=\frac{P_tA_t}{A_t^TP_tA_t+Q} |
---|
83 | \end{equation} |
---|
84 | \begin{equation} |
---|
85 | \hat{\theta}_{t+1}=\hat{\theta}_t+K_t(x_{t+1}-f_t(x_t,u_t)-A_t\hat{\theta}_t) |
---|
86 | \end{equation} |
---|
87 | \begin{equation} |
---|
88 | P_{t+1}=(I-K_tA_t)P_t |
---|
89 | \end{equation} |
---|
90 | |
---|
91 | \section{P�py k du�� �� |
---|
92 | nektere mozne pristupy, jak odhaduji suboptimalni $u_t$ |
---|
93 | \subsection{Certainty equivalecnce control} |
---|
94 | \subsection{Metoda separace} |
---|
95 | \subsection{SIDP} |
---|