2
AGGREGATION IN HIERARCHICAL SYSTEM: APPROXMATE DECOMPOSITIONS* Douglas P. Looze and Bassen Salhi Coordinated Science Laboratory University of Illinois Urbana, Illinois 61801 Abstract One problem facing hierarchical systems theory is that the allowed exact decompositions may result in problems which are too complex to be feasibly imple- mented. The approach of thispaper is toallowaggre- gations of the exact decomposition. The questions of convergence of the simplified iteration and the near- ness of the resulting solution to the true solution are addressed. 1. Introduction The topic of hierarchical control has-emerged as an important area in the study of large scale systems i n both theory and applications. Among t h ev a r i o u sa p p l i - cations which have been reported in the literature are applications to steel roughingprocesses[l],urban synchronousmachines[4],andriverpollutioncontrol traffic networks[2], water supplynetwork[31, class of problemswhichcanbehandled by current [5]. Despite this diversity of applications, the hierarchical techniques is limited. Several of the present limitations of hierarchical control theory are directly related to the technical development of the methodology which is in current use. The theory has developed as an extension of decomposi- tion techniques in mathematical programming to the particular structure of optimization problems considered by controltheorists. The theory insists on the decomposition of the original optimization problem with the iteration tending to the exact solution. This exactness requirement can lead to extremely complex subproblems i n many cases (e.g., stochastic problems or the completely decentralized linear quadratic problem). The purpose of this paper is to relax the exact- nessrequirement and defineapproximatedecompositions. The formulationofdecompositionalgorithmsdeveloped in [El is used and is restated in Section 2. In thisformulation,hierarchicaldecompositions are determined by one point iterations. In Section 3 sub- problem simplifications are described as operator theoretic aggregations of the iteration operator. The approximatedecomposition is then represented as a composition of an exact decomposition and an aggre- gation, and is defined in terms oftheaggregated iteration equation. Section 3also discussesrelationships between the original, exact decomposition and the resulting appro- nearly the simplified problem represents the original ximate decomposition. First,thequestionof how problem is addressed. Then, the convergence pro- perties of theapproximatedecomposition is related to those of the original decomposition. 2. Exact Decompositions For thepurposesofthispaper,thedecomposition formulation of 181 will be used. Assume the control problem to be solved can be expressed in the form f(x) = 0 (1) where 5 is a Banach space and f:Z+%. The desired *This work was supported in part by the Joint Services Electronics Program under Contract N00014-79- C-0424 and i n part by the National Science Foundation under Grant ENG-79-08778. solution of (1) will be denoted by x*. Let f :%x* +A be a continuously Frechet differentiable funceion such that a f (x*,x*) is invertible. Define the nonlinear split& f(x) = f0(X,Y) + fl(X,Y) VX,Y€%. (2) fo(%+l’%) + flf\,+ - 0. ( 3) The iteration defined by this splitting is given by the equation Conditions for local convergence and theasymptotic convergence rate of this iteration are given by the following theorem. Theorem 1: Assume f, fo, and flaredefined as above, and let x* be a solution to (1) for which af(x*) and alf(x*,x*) are nonsingular. If y= P~-alfo(X*,x*)-lalfl(x*,~*)}< 1 (4) then there exists an open neighborhood d Cs of x* such which satisfie; (3). Moreover that for any x €4 there is a unique sequence I%};-o lim “k= x* (5) k+= and for each E> 0 there exists an integer k, such that I\-x*l -< Vk_> ko. (6) Proof: See 181. A complete discussion of this formulation and its applicability to the analysis of hierarchical control algorithms can be found in (81 . It is sufficient to note here that mst suchalgorithmscanbedescribed and analyzed within this framework. above is that the analysis of general hierarchical structures is equivalent to the analysis of the non- linear discrete equation (3). In turn, properties of the iteration equation can be related to properties of the splitting function fo and the function f which describes the original problem. This advantage includes the ability to study the effects of problem simplifications at any level of the hierarchy. In particular, the simplification of these problems can be viewed as a nonlinear aggregation of (3). To illus- trate these ideas mre clearly, only the special case for which f and fo are linear in x will be considered in thispaper. The insights and results can be extended to the general nonliaear problem i n two ways: first throughthe use of linear aggregations at each iteration; or second, through the use ofanonlinear aggregation technique (c.f. is] 1. These cases will be considered in the full version of this paper. The advantage of the abstract formulation outlined Assume now that is a separable Hilbert space and thatequation(1) is linear in x Ax - b. (7) The linear splitting of the operator A is defined by a nonsingular operator A . :X+% and the equation Ax = Aox + %x. (8) The iteration equation can then be written A0\+li -9% + (9) The convergencecovdition(4) of Theorem 1 becows p{AilA., 1 1 (10) 445 0191-2216/80/0000-0445$00.75 0 1980 IEEE

[IEEE 1980 19th IEEE Conference on Decision and Control including the Symposium on Adaptive Processes - Albuquerque, NM, USA (1980.12.10-1980.12.12)] 1980 19th IEEE Conference on Decision

  • Upload
    hassen

  • View
    212

  • Download
    0

Embed Size (px)

Citation preview

Page 1: [IEEE 1980 19th IEEE Conference on Decision and Control including the Symposium on Adaptive Processes - Albuquerque, NM, USA (1980.12.10-1980.12.12)] 1980 19th IEEE Conference on Decision

AGGREGATION IN HIERARCHICAL SYSTEM: APPROXMATE DECOMPOSITIONS*

Douglas P. Looze and Bassen Sa lh i Coordinated Science Laboratory

Universi ty of I l l inois Urbana, I l l i n o i s 61801

Abstract

One problem facing hierarchical systems theory is that the allowed exact decompositions may result i n problems which a re too complex to be feas ib ly imple- mented. The approach of this paper is to allow aggre- gations of t he exact decomposition. The questions of convergence of the simplified i teration and the near- ness o f the resu l t ing so lu t ion to the t rue so lu t ion a re addressed.

1. Introduction

The topic of hierarchical control has-emerged as an important area i n the s tudy of large scale systems i n both theory and appl ica t ions . Among the var ious appl i - ca t ions which have been reported in t h e l i t e r a t u r e are app l i ca t ions t o steel roughing processes [l], urban

synchronous machines [4], and river pollution control t r a f f i c networks [2], water supply network [31,

c lass of problems which can be handled by current [5] . Despi te th i s d ivers i ty o f appl ica t ions , the

hierarchical techniques is l imi ted .

Several of the present l imitations of h i e ra rch ica l cont ro l theory a re d i rec t ly re la ted to the t echnica l development of the methodology which is in cu r ren t u se . The theory has developed as an extension of decomposi- t ion techniques in mathematical programming to the par t icular s t ructure of opt imizat ion problems considered by cont ro l theor i s t s . The theory insists on the decomposition of the original optimization problem with the i teration tending to the exact solution. This exactness requirement can lead to extremely complex subproblems i n many cases (e.g. , s tochas t i c problems or the completely decentralized l inear quadratic problem).

The purpose of t h i s paper is to re lax the exact- ness requirement and define approximate decompositions. The formulation of decomposition algorithms developed i n [ E l is used and is res ta ted in Sec t ion 2 . In this formulation, hierarchical decompositions are determined by one po in t i t e r a t ions . In Section 3 sub- problem s impl i f ica t ions a re descr ibed as operator theoret ic aggregat ions of t he i t e r a t ion ope ra to r . The approximate decomposition is then represented as a composition of an exact decomposition and an aggre- gation, and is defined in terms of the aggregated i te ra t ion equat ion .

Sec t ion 3a lso d i scussesre la t ionships between the original, exact decomposition and the resulting appro-

near ly the s implif ied problem represents the o r ig ina l ximate decomposition. F i r s t , the ques t ion of how

problem is addressed. Then, the convergence pro- pe r t i e s of the approximate decomposition is re la ted to those of the original decomposition.

2 . Exact Decompositions

For the purposes of this paper, the decomposition formulation of 181 w i l l be used. Assume the cont ro l problem to be solved can be expressed i n t h e form

f(x) = 0 (1)

where 5 is a Banach space and f : Z + % . The desired

*This work was supported i n p a r t by the Jo in t Services Electronics Program under Contract N00014-79- C-0424 and i n p a r t by the National Science Foundation under Grant ENG-79-08778.

solut ion of (1) will be denoted by x*. L e t f :%x* +A be a continuously Frechet differentiable funceion such t h a t a f (x*,x*) is invert ible . Define the nonl inear s p l i t &

f (x ) = f0(X,Y) + fl(X,Y) VX,Y€%. (2)

fo(%+l’%) + f l f \ , + - 0. ( 3)

The i te ra t ion def ined by t h i s s p l i t t i n g is given by the equation

Conditions for local convergence and the asymptotic convergence ra te of this i terat ion are given by the following theorem.

Theorem 1: Assume f , f o , and f l are def ined as above, and le t x* be a so lu t ion to (1) for which af(x*) and alf(x*,x*) are nonsingular. If

y = P~-alfo(X*,x*)-lalfl(x*,~*)}< 1 (4) then there exists an open neighborhood d Cs of x* such

which sa t i s f i e ; ( 3 ) . Moreover t h a t f o r any x € 4 there is a unique sequence I%};-o

lim “ k = x* (5) k+=

and for each E > 0 t h e r e e x i s t s an in teger k, such that

I\-x*l -< Vk_> ko. ( 6 )

Proof: See 181.

A complete discussion of this formulation and its appl icabi l i ty to the ana lys i s o f h ie rarch ica l cont ro l algorithms can be found i n (81 . It is s u f f i c i e n t t o no te here tha t mst such algorithms can be described and analyzed within this framework.

above is tha t t he ana lys i s of general hierarchical s t ruc tu res is equivalent to the analysis of the non- l inear d i scre te equa t ion ( 3 ) . In turn, propert ies of the i t e ra t ion equat ion can be re la ted to p roper t ies of the sp l i t t ing func t ion fo and the function f which descr ibes the o r ig ina l problem. This advantage includes the abi l i ty to s tudy the effects of problem s i m p l i f i c a t i o n s a t any l eve l of the hierarchy. In par t icu lar , the s impl i f ica t ion of these problems can be viewed as a nonlinear aggregation of ( 3 ) . To i l l u s - t r a t e t hese i deas mre c l ea r ly , on ly t he spec ia l ca se fo r which f and f o a r e l i n e a r i n x will be considered in this paper . The in s igh t s and results can be extended to the genera l nonl iaear problem i n two ways: f i r s t through the use of l inear aggrega t ions a t each i teration; or second, through the use of a nonlinear aggregation technique (c.f. is] 1. These cases w i l l be considered in the fu l l vers ion of th i s paper .

The advantage of the abstract formulation outl ined

Assume now tha t is a separable Hilber t space and that equation (1) is l i nea r in x

Ax - b. (7)

The l i n e a r s p l i t t i n g of the operator A is defined by a nonsingular operator A. : X + % and the equation

Ax = Aox + %x. (8)

The i te ra t ion equat ion can then be written

A0\+ l i -9% + ‘ (9) The convergence covdition (4) of Theorem 1 becows

p{AilA., 1 1 (10)

445 0191-2216/80/0000-0445$00.75 0 1980 IEEE

Page 2: [IEEE 1980 19th IEEE Conference on Decision and Control including the Symposium on Adaptive Processes - Albuquerque, NM, USA (1980.12.10-1980.12.12)] 1980 19th IEEE Conference on Decision

3. Approximate Decompositions

The d i f f icu l ty assoc ia ted wi th h ie rarch ica l decom- posit ion algorithms is t h a t t h e r e s u l t i n g i t e r a t i o n equation may be so complex that the implementation of the algorithm is not feasible . A natural approach to surmount t h i s d i f f i c u l t y is t o attempt to s impl i fy the i teration equation through an approximation to A,,. I n

on a subspace of the original space will be t h i s s e c t i o n , t h e problem of approximating an i t e r a t i o n

formulated.

L e t 3 be a sub-Iiilbert space o f 3 with the same topology +d l e t T :A+ 3 be a (possibly skew) projec- t ion onto&. The pseudo-inverse T i of T is defined uniquely by (91

T T T = T #

T'TT' = T

TT' = P (13)

T'T = I-Q (14)

where P and Q are the Or thogonal p ro jec t ions on to i and qT) r e spec t ive ly .

T onto X , the best approximation- of L on X defined by the p ro jec t ion T is given by

Giyen a l inear opera tor L :$+$, and-a pro jec t ion

L - T L T . # (15)

Thus, t h e i t e r a t i o n ( 9 ) r e s t r i c t e d t o 3 is given by

Two questions concerning the approximate i teration arise. F i r s t , what conditions can be placed on the subspzcez and the p ro jec t ion T such that the solut ion to (16) ( i f i t e x i s t s ) is nea r t he so lu t ion t o (7)? Secondly, what conditions can be placed o n 3 and T t o ensu re t ha t t he i t e r a t ion (16) converges?

Any s o l u t i o n t o (16) m u s t s a t i s f y i ; = i

where A = Ao+$ = TAT - b - #

GA= Tb. I f 3 is an invariant subspace of A (i.e., &) then the projecton onto 3 which minimizes the maximum e r r o r

e(T) = b sup IxCb)-T'i(b)l ( 2 0 ) I bl=1

is the orthogonal projection, and the value of (2O)_fs the spec t ra l rad ius of t h e o p e r a t o r r e s t r i c t e d t o 1. . I f b is f i x e d i n 4 t h e e r r o r ( 2 0 ) is zero for the orthogonal projection.

similar fashion. Assuming t h a t t h e o r i g i n a l i t e r a t i o n The convergence of (16) can be analyzed i n a

r e l a t e d t o t h a t of (9) through the concept of a (9) is convergent, the convergence rate of (16) can be

def la t ing pair of subspaces for (&,%) [lo]. A de f l a t ing pa i r of subspaces 74 and Y of J; are sub- spaces which are isomorphic and which s a t i s f y

AoZCClf A , U C t / . (21) This concept is re lated to the analysis of the conver- gence of (9) and (16) in the fo l lowing way.

The spec t r a l r ad ius of Ai1% is the magnitude of the largest generalized eigenvalue of (A,,,%). A gen- eral ized e igenvalue A of (Ao, and the corresponding generalized eigenvector x are 3' efined by the equation

AAox+qx = 0. (22)

Equation (22), %x, and A1x must l i e i n t h e same sub-

space i n which &x lies, a de f l a t ing pa i r by (211, gen- space. Thus, the subspace i n which x l i e s and the sub-

eral ize the concept of an invariant subspace to the generalized eigenvalue problem.

With th i s p re lude , ve can state the following theorem.

Theorem 21 L e t t 4 and 2/ be 5 def l a t ing pa i r of sub- spaces for (A,,,%) and le t y denote the magnitude of the largest associated generalized eigenvalue. Let T be a skew p r o j e c t i o n o n t o 4 a l o n g YL. Then the rate of convergence of (16) is c . Proof: L e t .

lJ = (U1,Uz) v = ( v V ) (23) 1' 2

be orthogonal operators with V1 :?/+x, v2 : f A + X , U1 : U + 3 , and U2 :2/'+%. Then, fo r i4,l

A, = V&.U*

us ing the par t i t ion def ined by (23). Thus, - Ai = TAiT #

But the propert ies of the skew pro jec t ion imply

m=(: 1) Hence (24) becomes

(25) . .

But (25) implies the generalized eigenvalues of (io,il) are those of (A&,<l) which i n t u r n are those of (&,A1) assoc ia ted wi th the def la t ing pa i r Y,V. .

1.

2.

3.

4.

5.

6.

7.

8.

9.

10.

References C. S. Eaglen, M. G. Singh, and J. F. Coales, "A Hierarchica l S t ra tem for the Temperature Control of a Hot Steel Rousing Process,"'Automatica, Vol. 9, pp. 209-222, 1973. M. G. Singh and 8. Tamura, "Modeling Hierarchical Optimization for Oversaturated Urban Road Tra f f i c Networks," In te rna t iona l J. of Control, Vo1..20,1974 F. Fa l l s ide and P. Perry. "Hierarchical Outimiza- t ion of a Water Supply Network," Proc. IEE,Vol. 122, 1975. M. Hassan and M. G. Singh, "The Hierarchical Control of a Synchronous Machine Using a Model Follower," Automatica, Vol. 13, March 1977. El. G. Singh and M. Hassan, "A Closed Loop Hierar- ch ica l Solu t ion for the Continuous T i m e River Pol lut ion Control Problem," Automatica, Vol. 12, May 1976. A. Benveniste, P . Bernhard, and G. Cohen, "On the Decomposition of Stochastic Control Problems," IFAC Symp. LSSTA, 1976. N. R. Sandell, Jr.. "Decomposition vs. Decentral-

~

ization in Large-Scale Systems Theory," 1979 IEEE CDC, 1976. D. P . Looze and N. R. Sandell, Jr., "Analysis of Decomposition Algorithms via Nonlinear Spl i t t ing Functions ," JOTA, July 1981 ( to appear). M. Z . Nashed, Ed., Generalized Inverses and Appli- ca t ions , Academic Press, New York, 1976. G. W. Stewart, "Error and Perturbation Bounds f o r Subspaces Associated with Certain Eigenvalue Problems," SIAM Review, Vol. 15, No. 4, 1973.

446