發(fā)布時間：2018-09-01 13:13

【摘要】：現(xiàn)實生活中,多數(shù)物理,醫(yī)藥,金融等問題均可由偏微分方程(PDEs)或者隨機偏微分方程(RPDEs)來描述.很多時候,人們不只關(guān)心PDEs或者RPDEs解本身的性質(zhì),更關(guān)心能否通過控制方程中的某些變量,使得另一些變量達到預期的狀態(tài),同時保證代價最小,這就是典型的PDE或者RPDE最優(yōu)控制問題.由于實際需求的驅(qū)動,系數(shù)確定和系數(shù)隨機情況下的PDE最優(yōu)控制問題得到了廣泛的研究和關(guān)注([44,56,58,131,144]).對于PDE最優(yōu)控制問題,其求解難點在于數(shù)值離散后,離散系統(tǒng)的規(guī)模較大,直接求解較困難,且在優(yōu)化過程中,需要求解大量代數(shù)方程組.對于RPDE最優(yōu)控制問題,它在PDE最優(yōu)控制問題的基礎(chǔ)上引入了隨機空間,因而問題更為復雜,一方面需要考慮適合于此問題結(jié)構(gòu)的隨機空間離散方法,另一方面原有定問題中的求解難點也因隨機空間的引入而增大.現(xiàn)有算法,對隨機空間離散后,多采用迭代法求解,在每步迭代中都需要求解大量定的PDEs,且內(nèi)存占用量很大.本文主要針對以上困難,提出了合理可行的數(shù)值算法,大大縮減求解方程的數(shù)量并降低了存儲代價.下面介紹本文中所用到的兩種核心數(shù)值方法.交替方向乘子法(Alternating Direction Method of Multipliers,ADMM),是一種求解結(jié)構(gòu)凸優(yōu)化問題的高效數(shù)值方法,主要利用分裂的思想,基于問題變量可分的特點,將大規(guī)模優(yōu)化問題分解成若干個規(guī)模較小的子問題.該方法具有全局收斂性,且最差情況下的收斂速度為O(1/k),其中k為迭代步數(shù).多重模式展開方法(Multi-Mode Expansion,MME),是求解隨機系數(shù)下RPDEs的有效算法.其主要思想是將隨機解按擾動量級的冪級數(shù)進行展開,展開系數(shù)滿足一系列迭代方程,并且迭代方程中系數(shù)為確定且相同的函數(shù),只有右端含有隨機項,從而降低了求剛度矩陣逆帶來的計算壓力.同時,該方法具有易模擬和高效的特點.本論文主要由兩部分構(gòu)成.首先,我們將ADMM應用到PDE最優(yōu)控制問題上,求解了幾類帶橢圓型PDE約束的最優(yōu)控制問題,給出完整的收斂性分析.其次,結(jié)合MME方法和Monte Carlo方法,我們將ADMM推廣到RPDE最優(yōu)控制問題上,在理論上證明了算法的收斂性,在數(shù)值上驗證了算法的有效性.具體工作如下:第一部分是本文的第二章,以Poisson方程為例,研究帶橢圓型PDE約束的最優(yōu)控制問題.主要考慮了在無約束和盒子約束下,控制變量分別為分布控制,Dirichlet邊界控制和Robin邊界控制等六種情況.問題描述為其中,針對分布控制問題和邊界控制問題,積分區(qū)間B分別取做D或(?)D,相應地,dz取做dx或ds.這里,控制約束條件分別取Uad=U(無約束)和Uad={u(a)∈U|ua≤u(x)≤ub}(盒子約束)兩種不同情況.狀態(tài)方程e(y,u)= 0分別取做以下三類控制問題對應的變分形式:本文首先考慮對上述問題進行有限元離散,并將三類狀態(tài)方程的離散格式寫成統(tǒng)一的線性方程組形式.至此,模型問題化成了有限維的優(yōu)化問題,形式如下:接著,我們分別針對無控制約束條件和盒子控制約束條件,采用ADMM和two-block ADMM求解上述離散問題.與現(xiàn)有算法相比,本文所提算法的優(yōu)勢在于避免了求解大規(guī)模離散系統(tǒng),且在每步迭代時無需求解方程組,只需在迭代循環(huán)外,求兩次系數(shù)矩陣的逆即可.此外,本文還給出了所提算法的收斂性分析,結(jié)果如下:定理 設(shè)h為有限元方法網(wǎng)格步長,k為ADMM的迭代步數(shù).令u*和uhk分別表示原問題的解和ADMM求出的迭代解.則有誤差估計‖u*-Ru_h~k‖_(L~2(D))=O(hp)+O(k~(-1/2)),成立,其中p為常數(shù)(對不同的模型取不同值,參見論文第二章).最后,數(shù)值實驗表明,本文所提算法能夠有效地處理PDE最優(yōu)控制問題.第二部分由本文的三,四章組成,我們將ADMM推廣到了如下的RPDE最優(yōu)控制問題在第三章中,上述模型中的狀態(tài)方程s(ξ,y(x,ξ),u(x))=0取為如下的隨機Poisson方程,基于Monte Carlo方法和有限元離散,容易得到上述模型的離散優(yōu)化問題為:傳統(tǒng)的方法(例如采用Monte Carlo方法離散隨機空間,并結(jié)合Newton迭代法或SQP等算法求解離散后的優(yōu)化問題),在計算過程中各個樣本之間會耦合在一起形成一個超大規(guī)模系統(tǒng).在本章里,我們首先用Monte Carlo方法結(jié)合有限元方法,給出問題的離散格式.隨后,根據(jù)所得離散問題的全局一致特性,采用ADMM分裂求解.迭代過程中,每個樣本都對應一個與定的PDE最優(yōu)控制問題同規(guī)模的子問題,且基于不同樣本,我們可以采用并行計算.對比來說,本章所提算法的優(yōu)勢依然在于ADMM,避免求解大規(guī)模離散系統(tǒng).只需在迭代過程中對于每個樣本求解相應的低維子問題.特別地,全程不涉及求解PDEs,只需在ADMM迭代循環(huán)外,對每個樣本點求系數(shù)矩陣的逆,保存下來,迭代過程中直接使用.本文還給出了算法的整體誤差估計,結(jié)果如下:定理 設(shè)M為采用的Monte Carlo樣本數(shù),h為有限元方法網(wǎng)格步長,k為ADMM迭代步數(shù).令u*和u≤分別為原問題的解和ADMM迭代解,則有如下離散誤差估計‖u*-Ruhk‖L2(D)≤O(M1/2)+O(h2)+O(k-1/2),在第四章中,我們繼續(xù)研究RPDE最優(yōu)控制問題.主要考慮狀態(tài)方程s(ξ,y(x,ξ).u(x))= 0取為如下的隨機Helmholtz方程.對于該問題,本章采取了三種算法來求解.第一種算法是直接將第三章的方法,平行推廣到帶隨機Helmholtz方程約束的最優(yōu)控制問題.由于本章模型問題需要在復空間中求解,且對網(wǎng)格精度有一定要求,故算法1存在一定的缺陷:(1).需要求解M+1次系數(shù)矩陣的逆,M為樣本數(shù);(2).算法內(nèi)存花費量為O(MN~2),N為物理空間自由度.當M和N很大時,算法需要的計算量和內(nèi)存量是難以忍受的.值得一提的是,對于求解RPDE最優(yōu)控制問題的一般方法,如Stochastic Collocation結(jié)合CG法,Monte Carlo方法結(jié)合SQP方法等,也存在上述兩個問題.針對以上兩個難點,本文采用MME方法,對最優(yōu)控制問題進行預處理.在算法2中,我們基于MME,有限元離散和Monte Carlo方法,將原問題轉(zhuǎn)化為一個離散的最優(yōu)控制問題,然后利用ADMM進行迭代求解.這種方法全程只需要求解兩次系數(shù)矩陣的逆;在計算量上已經(jīng)較一般方法有了顯著的優(yōu)勢,但此算法的內(nèi)存花費量依然為O(MN~2).接著,我們依據(jù)MME的一些迭代特性,提出了一個重要的等式關(guān)系,并利用此等式關(guān)系,在算法2的基礎(chǔ)上設(shè)計了算法3.此方法能使得RPDE最優(yōu)控制問題中的隨機域全部轉(zhuǎn)移到目標泛函帶有期望的系數(shù)上,使隨機更容易處理.經(jīng)過簡單計算后,原隨機問題可變?yōu)槎ǖ腜DE最優(yōu)控制問題,從而避免了求解隨機最優(yōu)控制問題時會遇到的困難.該算法全程只需求兩次系數(shù)矩陣的逆,無需額外求解方程組,且內(nèi)存花費量僅為O(N~2),與M無關(guān),從本質(zhì)上解決了上述兩個難點,是三種算法中效果最好的.算法3的收斂階估計如下:定理設(shè)M為利用Monte Carlo方法的樣本數(shù),ε為隨機折射率中的擾動量級,Q為MME方法的展開項數(shù),h為有限元方法的網(wǎng)格步長,k為迭代步數(shù).則原問題最優(yōu)解u*和本文所提算法得到的數(shù)值解uhQ,k間的誤差估計為最后,我們通過數(shù)值模擬驗證了本文所提算法的高效性.
[Abstract]:In real life, most physical, medical and financial problems can be described by partial differential equations (PDEs) or stochastic partial differential equations (RPDEs). In many cases, people are not only concerned about the properties of PDEs or RPDEs solutions, but also about whether some variables in the governing equations can make other variables reach the desired state, while ensuring that This is a typical PDE or RPDE optimal control problem with minimal cost. Driven by actual demand, the PDE optimal control problem with deterministic coefficients and stochastic coefficients has attracted extensive research and attention ([44,56,58,131,144]). It is difficult to solve directly and a large number of algebraic equations need to be solved in the process of optimization. For RPDE optimal control problem, stochastic space is introduced on the basis of PDE optimal control problem, so the problem is more complicated. On the one hand, the stochastic space discretization method suitable for the structure of the problem should be considered, on the other hand, the original definite problem should be solved. The difficulty of solving the problem is also increased by the introduction of the random space. The existing algorithms, after the discretization of the random space, usually use iterative method to solve the problem. In each iteration, they need to solve a large number of PDEs, and the amount of memory occupied is very large. In this paper, two core numerical methods are introduced. Alternating Direction Method of Multipliers (ADMM) is an efficient numerical method for solving structural convex optimization problems. Based on the idea of splitting, large-scale optimization problems are decomposed into discrete variables. Several small-scale sub-problems. The method has global convergence and the worst-case convergence rate is O(1/k), where k is the iterative step. Multimode Expansion (MME) is an effective algorithm for solving RPDEs with random coefficients. The expansion coefficients satisfy a series of iteration equations, and the coefficients in the iteration equations are deterministic and the same function. Only the right end of the iteration equation contains random terms, which reduces the computational pressure caused by the inverse of stiffness matrix. In the optimal control problem, several kinds of optimal control problems with elliptic PDE constraints are solved and the complete convergence analysis is given. Secondly, combining MME method with Monte Carlo method, we extend ADMM to RPDE optimal control problem. The convergence of the algorithm is proved theoretically and the effectiveness of the algorithm is verified numerically. The first part is the second chapter of this paper. Taking Poisson equation as an example, the optimal control problem with elliptic PDE constraints is studied. The control variables are distributed control, Dirichlet boundary control and Robin boundary control respectively under unconstrained and box constraints. The problem is described as distributed control problem and boundary control problem. For control problems, the integral interval B is taken as D or (?) D respectively, and DZ is taken as DX or ds. Here, the control constraints are taken as two different cases: Uad = U (unconstrained) and Uad = {u (a) - {u | u a U < u (x) < {UB box constraints. The state equation E (y, u) = 0 is taken as the corresponding variational form of the following three types of control problems: Firstly, the above-mentioned question is considered. The problem is discretized by finite element method, and the discrete schemes of three kinds of state equations are written into a unified linear system of equations. Thus, the model problem is transformed into a finite-dimensional optimization problem in the following forms: Then, for uncontrolled constraints and box control constraints, we use ADMM and two-block ADMM to solve the above discrete problems respectively. Compared with other algorithms, the advantage of the proposed algorithm is that it avoids solving large-scale discrete systems and does not need to solve equations at each iteration. It only needs to find the inverse of the coefficient matrix twice outside the iteration cycle. K is the iteration step of ADMM. Let u* and u H K represent the solution of the original problem and the iterative solution of ADMM, respectively. Then the error estimate _*-Ru_h~k_ (L~2(D))= O(h p) + O (k~(-1/2)) holds, where p is a constant (for different models, see Chapter 2 of this paper). Finally, numerical experiments show that the proposed algorithm can deal with PDE most effectively. The second part is composed of three and four chapters of this paper. We generalize ADMM to the following RPDE optimal control problems. In the third chapter, the state equation s (zeta, y (x, zeta), u (x))= 0 in the above model is taken as the following stochastic Poisson equation. Based on Monte Carlo method and finite element discretization, the discrete optimization problem of the above model can be easily obtained. The Title is: Traditional methods (such as Monte Carlo method to discretize random space and Newton iterative method or SQP algorithm to solve discrete optimization problems) will be coupled to form a very large-scale system. In this chapter, we first use Monte Carlo method and finite element method to solve the discrete optimization problems. Then, according to the global uniformity of the discrete problem, the ADMM splitting method is used to solve the problem. Each sample corresponds to a sub-problem of the same size as the fixed PDE optimal control problem in the iterative process, and based on different samples, we can adopt parallel computing. The advantages of the proposed algorithm in this chapter are still in comparison. In ADMM, the solution of large-scale discrete systems is avoided. Only the corresponding low-dimensional sub-problems are solved for each sample in the iterative process. Particularly, the whole process does not involve solving PDEs, only the inverse of the coefficient matrix is obtained for each sample point outside the iterative cycle of ADMM, which is saved and used directly in the iterative process. The results are as follows: Theorem M is the number of Monte Carlo samples used, h is the mesh step of the finite element method, and K is the ADMM iterative step. If U * and U < are the solution of the original problem and the ADDM iterative solution respectively, then the following discrete error estimates -Ruhk L2 (D) O (M1/2) + O (h2) + O (k-1/2) are obtained. In Chapter 4, we continue to study RPDE optimal control. This paper mainly considers the stochastic Helmholtz equation in which the state equation s (zeta, y (x, zeta).U (x) = 0 is taken as the following. Three algorithms are used to solve this problem. The first algorithm is to extend the method in Chapter 3 directly to the optimal control problem with stochastic Helmholtz equation constraints. Because the model problem in this chapter needs to be solved in complex space. It needs to solve the inverse of M+1 coefficient matrix and M is the number of samples; (2) The memory cost of the algorithm is O (MN~2), and N is the degree of freedom of physical space. When M and N are very large, the amount of computation and memory required by the algorithm is unbearable. The general methods of DE optimal control problems, such as Stochastic Collocation combined with CG method, Monte Carlo method combined with SQP method, also have the above two problems. To solve the above two difficulties, MME method is used to preprocess the optimal control problem. In algorithm 2, based on MME, finite element discretization and Monte Carlo method, the original problem is solved. This method only needs to solve the inverse of the coefficient matrix twice in the whole process. It has a remarkable advantage over the general method in computation, but the memory cost of this algorithm is still O (MN~2). Then, according to some iterative characteristics of MME, we propose a new method. Using this equation, an algorithm is designed on the basis of algorithm 2. This method can transfer all the random fields in the RPDE optimal control problem to the target functional with expected coefficients, making the stochastic process easier. After simple calculation, the original stochastic problem can be changed into a PDE optimal control problem. The algorithm only needs the inverse of the coefficient matrix twice in the whole process, and does not need to solve the equations. The memory cost is only O (N~2), and it is independent of M. It essentially solves the above two difficulties and is the best of the three algorithms. In order to use the sample number of Monte Carlo method, e is the order of perturbation in random refractive index, Q is the expansion term of MME method, h is the mesh step of finite element method and K is the iteration step, then the error estimation between the optimal solution u* and the numerical solution u h Q, K obtained by the proposed algorithm is finally verified by numerical simulation. The efficiency of the algorithm.
【學位授予單位】：吉林大學
【學位級別】：博士
【學位授予年份】：2017
【分類號】：O241.82;O232

【相似文獻】

相關(guān)期刊論文 前10條

1 羅金火,潘立平;終端受限的線性-非二次最優(yōu)控制問題[J];復旦學報(自然科學版);2003年02期

2 邢進生,劉人境,李晉玲;一個有效兩階段最優(yōu)控制問題的算法[J];北京電子科技學院學報;2004年04期

3 佟欣;張洪光;;一類生態(tài)系統(tǒng)的最優(yōu)控制問題[J];生物數(shù)學學報;2013年03期

4 俞玉森;評《最優(yōu)控制問題的計算方法》[J];數(shù)學研究與評論;1981年S1期

5 吳鐵軍,呂勇哉;一種求解帶約束最優(yōu)控制問題的算法[J];控制理論與應用;1986年04期

6 卪亮壯;醫(yī)學中的一個最優(yōu)控制問題[J];北京航空學院學報;1988年03期

7 趙寶元;氣-固反應中的一個最優(yōu)控制問題[J];高校應用數(shù)學學報A輯(中文版);1990年02期

8 王玲,李建國,斯洛齊克;解決最優(yōu)控制問題的準梯度方法(英文)[J];控制理論與應用;1999年03期

9 楊然,周鋼,許曉鳴;求解最優(yōu)控制問題的改進辛幾何算法[J];上海交通大學學報;2000年04期

10 楊然,周鋼,許曉鳴;求解最優(yōu)控制問題的改進辛幾何算法[J];上海交通大學學報;2000年05期

相關(guān)會議論文 前10條

1 潘立平;周淵;;線性非二次最優(yōu)控制問題的一種解法[A];第二十七屆中國控制會議論文集[C];2008年

2 張寶琳;樊銘渠;;一類奇異時滯系統(tǒng)奇異二次指標最優(yōu)控制問題的近似方法[A];第二十七屆中國控制會議論文集[C];2008年

3 李春發(fā);陳華;;古地溫度場系統(tǒng)的參數(shù)識別及最優(yōu)控制問題[A];中國運籌學會第六屆學術(shù)交流會論文集（上卷）[C];2000年

4 高彩霞;馮恩民;;一類以脈沖系統(tǒng)為約束最優(yōu)控制問題的優(yōu)化算法[A];中國運籌學會第八屆學術(shù)交流會論文集[C];2006年

5 唐萬生;李光泉;;時變廣義系統(tǒng)最優(yōu)控制問題[A];全國青年管理科學與系統(tǒng)科學論文集（第1卷）[C];1991年

6 雍炯敏;;具有狀態(tài)約束的二階半線性橢圓型方程的最優(yōu)控制問題[A];1991年控制理論及其應用年會論文集（下）[C];1991年

7 肖華;吳臻;;一類線性二次正倒向隨機控制系統(tǒng)的最優(yōu)控制問題[A];第二十三屆中國控制會議論文集（上冊）[C];2004年

8 陶世明;朱經(jīng)浩;;Canonical對偶方法與一類最優(yōu)控制問題[A];中國運籌學會第九屆學術(shù)交流會論文集[C];2008年

9 楊富文;;求一類H~∞最優(yōu)控制問題的非迭代算法[A];1992年中國控制與決策學術(shù)年會論文集[C];1992年

10 王水;朱經(jīng)浩;;線性規(guī)劃在半定二次最優(yōu)控制問題中的應用[A];中國運籌學會第八屆學術(shù)交流會論文集[C];2006年

相關(guān)博士學位論文 前10條

1 李景詩;PDE和RPDE最優(yōu)控制問題的交替方向乘子法[D];吉林大學;2017年

2 邵殿國;若干正倒向隨機比例系統(tǒng)的最優(yōu)控制問題[D];吉林大學;2015年

3 鞏本學;具有隨機場系數(shù)偏微分方程的最優(yōu)控制問題數(shù)值方法[D];山東大學;2016年

4 王海洋;時間不相容的隨機控制問題和弱形式的正倒向隨機微分方程[D];山東大學;2016年

5 張倩;幾類PDE約束最優(yōu)控制問題的數(shù)值方法研究[D];南京師范大學;2016年

6 劉平;控制變量參數(shù)化最優(yōu)控制問題計算方法研究[D];浙江大學;2017年

7 張穩(wěn);若干微分方程最優(yōu)控制問題的譜方法[D];上海大學;2009年

8 郭磊;混合動態(tài)系統(tǒng)建模、穩(wěn)定性及最優(yōu)控制問題研究[D];山東大學;2006年

9 李彬;含狀態(tài)和控制約束的最優(yōu)控制問題和應用[D];哈爾濱工業(yè)大學;2011年

10 唐躍龍;兩類最優(yōu)控制問題變分離散方法的研究[D];湘潭大學;2012年

相關(guān)碩士學位論文 前10條

1 張培勇;時標上一類最優(yōu)控制問題研究[D];貴州大學;2009年

2 管文君;發(fā)展方程的能控性和最優(yōu)控制問題[D];東北師范大學;2015年

3 黃啟燦;數(shù)值天氣預報模式誤差項的最優(yōu)控制問題研究[D];蘭州大學;2015年

4 方研;帶有終端角度和攻擊時間約束的協(xié)同制導律設(shè)計[D];哈爾濱工業(yè)大學;2015年

5 夏云飛;一類滿足Lotka-Volterra互惠關(guān)系的生物種群最優(yōu)控制問題[D];哈爾濱師范大學;2015年

6 邵志政;帶有非線性干擾補償?shù)腁DP控制方法及在風機變槳控制的應用[D];東北大學;2014年

7 李越;基于空間分數(shù)階擴散方程及點態(tài)受限約束的三維最優(yōu)控制問題的快速算法[D];山東大學;2016年

8 孫肖斌;帶擴散的對偶模型的最優(yōu)分紅與注資[D];曲阜師范大學;2016年

9 劉志博;Navier-Stokes方程約束最優(yōu)控制問題的分裂預處理迭代方法[D];南京師范大學;2016年

10 蔡超;三類發(fā)展型方程的系數(shù)反演問題[D];蘭州交通大學;2016年

，

本文編號：2217309