For later use we dene a constraint-set Z k utilizing an expression for the state predictor Z k (x k )=fz k =( k k ): k 2 ^ k 2 ^ k = F k ( k x k )g 8

Transcription

1 IDENTIFICATION AND CONVEITY IN OPTIMIZING CONTROL Bjarne A. Foss Department of Engineering Cybernetics, Norwegian University of Science and Technology, N-734 Trondheim, Norway. Tor A. Johansen SINTEF Electronics and Cybernetics, Automatic Control Department, N-734 Trondheim, Norway. Abstract. The interplay between identication of a predictor and the convexity of an associated optimization-based control problem is studied. The reason for the appearance of nonconvexity is investigated and a method to favour convexity of the optimization-based control problem in the identication problem is suggested. Further, a parametrization which makes the identication problem tractable is proposed. Simulation results suggest a merit to method. Keywords. Optimization, nonconvex problems, identication, nonlinear systems, model predictive control 1. INTRODUCTION The combined use of dynamic models and optimization oers a concept in which process knowledge can be linked to operational goals formulated by some optimization criterion. This concept has seen widespread use. One such area is process control and the use of model predictive control (MPC). MPC refers to a class of algorithms where an optimization problem is solved repetitively, at every new time-instant, see for example Lee (1996). The interaction between control and optimization is discussed in an illuminating way in Mayne (199). In this paper the divide between convex and nonconvex problems is emphasized. In particular problems arise by using a nonlinear predictor as opposed to a linear predictor since the optimization problem in the former cases in general becomes nonconvex. This latter observation forms the departure point for this study. The paper is structured as follows: First, a problem denition is provided. Thereafter, we investigate convexity in nonlinear optimization-based control. In particular, we discuss design tradeos between identifying a predictor, and convexity of an associated optimization problem. Section 4 highlights some of the ndings by a simulation example, while a discussion and some conclusions ends the paper. 2. PROBLEM DEFINITION The optimality criterion is dened in discrete-time k (z k )= k ( k k )= i2i k l i (x i+1 u i ) (1) where k = fu T k ::: ut k+n;1 gt =U ::: U k = fx T k+1 ::: x T N g T = ::: I k = fk : : : k + N ; 1g u i 2 U R mu x i 2 R mx u i denotes the control input of a system. Since we are assuming a time-discrete model the control input u i is constant during the time span [i i + 1 >. x i denotes the system state at time i. Above, the sets and (or U and ) invoke constraints on k and k. The states and control inputs are, however, not independent, since the state sequence k in a dynamic system is uniquely dened by the initial value, ie. x k, and the control input sequence k.we dene a state-space model, valid on at least I k, and a predictor by h i (x i+1 x i u i )=x i+1 ; f i (x i u i )= (2) k = F k ( k x k ) (3) = ff k1 ( k x k ) T ::: F kn ( k x k ) T g T which means that F k performs the mapping F k :!. The predictor can be realized by multiple use of a state-space model. The minimization problem can be stated as k o = arg min k 2 k ( k k ) (4) subject to k = F k ( k x k ) 2 x k ; given 1

2 For later use we dene a constraint-set Z k utilizing an expression for the state predictor Z k (x k )=fz k =( k k ): k 2 ^ k 2 ^ k = F k ( k x k )g 8 x k 2 () 3. CONVEITY ISSUES IN OPTIMIZATION BASED CONTROL We will in the following categorize optimization problems arising in control and consider the tradeos which should be made. First, we separate the convex from the nonconvex cases. 3.1 Convex problems We rst note that problem (4) is convex if k is a convex functional and Z k is a (nonempty) convex set. The issue of k being a convex functional does not impose severe restrictions in optimizing control. The optimization criterion will typically be de- ned by some distance measure based on a norm, and any norm is a convex functional. The problem of limiting Z k (x k ) 8 x k 2 to a convex set imposes severe restrictions from a control point of view. To discuss this we rst derive sucient conditions for Z k (x k ) to be a convex set. Sucient conditions for convexity The set Z k (x k ) in () is convex if the following assumptions are satised. A1: is a convex set. A2: is a convex set. A3: The mapping F k ( x k ):! R mx :::R mx is ane for all x k 2. A4: Z k (x k ) is nonempty for all x k 2. The proof is omitted here due to page limitations. The conditions are sucient, since there exists nonane mappings F k that renders a convex set Fk (x k ). This is readily seen in the 1-dimensional case where any continuous mapping provides a convex set Fk (x k ). We will in the following discuss assumptions A1 to A4 above with particular emphasis on how they restrict the control problem formulation. Since is convex the set U must be convex, cf. (4). The control input at a given time instant is chosen from U, ie. u i 2 U R mu. If the restrictions on the m u dierent control inputs are independent U will be convex if each of the control inputs are continuous in the sense that u ij 2 [u j u j ] R 8j 2 f1 ::: m u g. Dening U by including dependency between the control inputs will in most cases not introduce problems. Typical limitations like ju ij j + ju il j1, ju ij + u il j1, or ku ij + u il k 2 1, where u ij and u il denote two dierent control inputs at a time i, do not by themselves impose nonconvexity. As a conclusion Assumption A1 does not impose severe restriction on the formulation of the control problem as long as the control inputs are continuous. The discussion of Assumption A2 is very similar to the discussion of A1 above. Hence, if the states are continuous A2 does not impose severe restriction on the problem formulation. A3 is the problematic assumption since it severely limits the class of possible predictors. Viewing linear predictors, they always satisfy this assumption. Nonlinear predictors also satisfy this condition as long as they remain linear in the control input k. Assumption A4 is related to feasibility, ie. is it possible to nd a control sequence that satises the state constraint?we will assume that this assumption always is satis- ed. From a control design point of view this should be interpreted as a controllability constraint. 3.2 Nonconvex problems We will now discuss nonconvex problems, and assume that Assumptions A1, A2 and A4 are satis- ed. This means that nonconvexity is caused by the predictor alone. By the former discussion we claim that this is an important cause for nonconvexity, hence this type of nonconvex problems is highly relevant to optimization-based control Identication method Let us rst formalize an identication problem tailored for optimization-based control. The issue is to construct a predictor which produces small prediction errors on the optimization horizon, eg. I k in (1). First we dene a set of predictors parameterized by some parameter, cf. (3). F = ff ( ) : 2 g R m (6) For the sake of convenience and notation we have limited the set of predictors above in the sense that they are time-invariant as compared to the time-varying predictors in (3). We assume to have a set of input and output data for the system in question and structure these in an information vector. The output is assumed to be a function of the current state. Further, we also dene a truncated information vector.

3 I M = fu y 1 ::: u M;1 y M g (7) y i 2 Y R my y i = g(x i ) I i = fu y 1 ::: u i;1 y i g (8) This allows us to dene a causal predictor for the system output. ^y i+jji = g(^x i+jji )=g(f j ( i ^x iji )) (9) 8i 2 I j 2 I N ^y i+jji ^x i+jji ; predicted output or state given I i ^x iji ; state estimate of x i byuseofi i I = f ::: M ; Ng j 2 I N = f ::: Ng F j refers to the j-step ahead predictor, ie. the j'th componentoff j dened in (3). Note that the predictor is time-invariant. We observe that the prediction is based on a current state estimate ^x iji and the use of the predictor F j. In the following we will make the following two simplications. We will assume that the state is known in the sense that ^x iji = x i. Hence, we do not address the state-estimation problem. We further assume that the system output function g is known which means that the system output is known, ie. ^y iji = y i. The identication problem is posed as o = arg min 2 V () (1) V ( I M F g)= l pj (y i+j ^y i+jji )(11) i2i j2i N l pj : Y Y! R Let us consider the inner part of the optimality criterion. At a given time step i 2 I, the inner sum computes the eect of the prediction error on the horizon fi+1 ::: i+ng. l pj is some metric. It may vary on the prediction horizon. In particular, it can be of interest to fade away the eect of prediction errors for large values of j. Ultimately, we would like to identify predictors which render a convex optimization problem (4). Two means of attacking this problem are (i) limiting the class of predictors F, and (ii) modifying modify the identication criterion (11). Limiting the class of predictors to comply with convexity leaves us with ane predictors, or even more limited, linear predictors. Viewing the identication problem the minimum criterion value V o = V ( o I M F g) will increase if the size of the set of predictor F decreases. Hence, there is a trade-o between the quality of the predictions and the advantage of having a convex optimization problem. Referring to (ii), limiting the class of predictor can be viewed as inferring hard constraints on the optimization problem. An alternative is to include soft constraints by modifying the identication criterion in (1). In this work we do this by using ideas from (Johansen 1996) in which the identication criterion is modied to account for prior knowledge or desired properties on a problem. The modied identication criterion is given by: V ( I M F g)= i2i i2i N l pj (y i+j ^y i+jji ) + ks F F k + kf ; F lin k (12) 2 [ 1) S F is an operator that projects the nonlinearities of F onto a normed function space. S F will typically contain 2nd order derivative terms which vanish if F is linear. kf ; F lin k is a distance measure between an identied predictor F and the best linear predictor, F lin. The norm is typically a weighted 2-norm. Let us view two possibilities. If > = the predictor nonlinearities are limited by a trade-o between the prediction error and the degree of nonlinearity of F. Second, if = > it is necessary to obtain an apriori linear predictor. The distance from this linear model is then penalized. A problem with the aboveapproach is to compute the norms of S F F or F ; F lin, in particular their gradients with respect to. It is highly desirable to have a linear parameterization since this greatly reduces the computational eort. The proposed method may imply a signicant complexity increase compared to the standard identication formulation. The identication problem is, however, an o-line problem. Further, the complexity increase is motivated by the fact that we are seeking predictors that render a convex or close-to-convex optimization-based control problem. These are on-line problems in which timeecient algorithms can be critical for stability and performance. The method adds soft constraints in the sense that an identied predictor may well produce a nonconvex problem. We do, however, know that a linear predictor renders a convex problem given Assumptions A1, A2 and A4. Hence, if the deviation from a linear predictor is penalized it seems reasonable to believe that this convexies the optimization problem. Alternative methods can be suggested by including convexity conditions as hard constraints in the identication problem.

4 3.3 Convexity measure To the authors' knowledge there does not exist any measures of the degree of convexity in the literature. To evaluate our results in quantitative terms, however, we dene a convexity measure based on the epigraph of over, cf. (Luenberger 1969), p.192. We dene the set (v x k )=f 2 :( k ) v (13) ^ k = F k ( x k ) 2g v 2 R Further we dene the convex hull of following Floudas (Floudas 199), p.21. Denition of a convexity measure Let (v x k ) be a set (convex or nonconvex) in R m. The convex hull, denoted ~(v x k ), of (v x k ) is the intersection of all convex sets in R m which contains (v x k ) as a subset. Hence, the convex hull is the smallest extension ~(v x k )of(v x k )such that ~(v x k )isconvex. If (v x k ) is a convex set, it is equal to its convex hull. The convexity measure is based on the volume (v x k ) compared to its convex hull. = 1 v ; v Z v v R ~(v x k );(v x k ) d R ~ (v xk ) d dv (14) By choosing v equal to the global minimum of ( k ) the denominator will always be larger than. v is given by subject to v = max k 2 k ( k k ) (1) k 2 h i (x i+1 x i u i )= 8 i 2 I k x k ; given Because of the normalization the convexity measure is bounded by and 1, ie. 2 [ 1). The convexity measure will be computed by discretizing uniformly, and discretizing v. The computation is extensive, in particular the calculation of the convex hull. 3.4 Parametrization The identication problems posed using V can easily become intractable. It is particularly important tochoose a parametrization which limits the complexity. There are multiple choices. In this work we choose a very simple parametrization. We assume that a linear predictor and a nonlinear predictor has been identied by means of (1). A new predictor is dened as a convex combination of these two predictors. F = p F l +(1; p )F n p 2 [ 1] (16) F l and F n are the linear and nonlinear predictors identied by (1). Hence, one parameter, p, the weighting between the linear and nonlinear predictor, needs to be estimated. p will take onvalues close to if the inuence of the terms ks F F k and kf ;F l k is small compared to the prediction error term, provided F n is a better predictor than F l. 4. EAMPLE In this example we investigate properties related to a model of a continuously stirred tank reactor with the following three reactions: A! k1 B! k2 C 2A k3x3! D E g(u)! 2E, and state-space model formulated in continuous time. _x 1 = ;k 1 (u)x 1 ; k 3 (u)x 3 x v V (x 1 ; x 1 ) _x 2 = k 1 (u)x 1 ; k 2 (u)x 2 ; v V x 2 _x 3 = g(u)x 3 + v V (x 3 ; x 3 ) (17) y = x 2 k i (u)=k i e ; E i u i 2f1 2 3g g(u)=be ;(u;us)2 Parameter values are given in Table 1. k 1 k 2 k 3 1:287e +12, 1:287e + 12, 9:43e +9 E 1 E 2 E 3 978:3, 978:3, 86: us 387:3K v=v 14:18 b 1: x 1 x 3 :1 1: Table 1. Example 2: Parameter values. x 1 x 2 x 3 u are the concentrations of A, B and E, and temperature, respectively, while y constitutes the process output. E is some biochemical substance which grows according to temperature conditions given by g(u). It uses some inert substance to support this growth. E catalysis the reaction 2A! D. We assume that B is the valuable product and hence try to maximize the productionofthisover some time-horizon. If the reaction E! 2E is removed the model resembles a Van de Vusse reactor. We dene the following criterion using a sampling time of :1. ( )=; i=1 jy i j (18) = fu u u u 1 u 1 u 1 u 1 u 1 u 1 u 1 g = fx 1 ::: x 1 g U =[u s ; : u s +:] =[1: 2:] [: 2:] [:8 2:] x ; given

5 The graph of, ie. a surface in R, using the state-space model (17) for prediction, is shown in the upper left part of Fig.3 for dierent control inputs and initial conditions x =(2:143 1:9 1:). Nonconvexity is readily observed in this case. The nonconvexity is mainly caused by g(u). The global minimum for is obtained for (u u 1 )=(2:1 :). The surface changes with dierent values x 2. The nonconvexity is, however, maintained. We will assume that the nonlinear process model is known. This forms the basis for the nonlinear predictor. This means that the optimal control input, in this case, is searched for on a graph like the upper left part of Fig.4. A linear predictor is identied on the basis of the two solid line input data sequences shown in Fig.1 and the two corresponding output sequences shown in Fig.2. This gives 2 data points. The initial condition is x =(2:143 1:9 1:) T. A constant level is removed from each of these data sequences. No noise is added to the data. The identication criterion is identical to (11) with l pj as a quadratic function, ^y iji = y i and I N = f1 ::: 1g. The linear predictor is based on multiple use of the 1-step linear predictor. We use the Akaike Final Prediction Error to choose the model structure. y i+1 =1:3y i ; :48y i;1 (19) +:9u i ; :68u i;1 Since the model is based on data in which the bias is removed the input and output in (19) must be adjusted when used for prediction. In practice the input/output data to (19) can be corrected by subtracting a lowpass-ltered value of the input and output. This is done in the prediction evaluation below. To illustrate the prediction properties of the linear model, the 1-step ahead prediction, ie. :1 time units, is shown in Fig.3 for the initial condition x =(2:1 1:84 1:2) and a control input sequence dierent from the identication data set, see the validation data in Fig.1. The prediction is quite good. It should be remarked that if the control input gets close to u s the dynamics change signicantly. No identication data for control inputs in this region is available. Hence, the prediction using the linear model will deteriorate significantly for validation data in this region. The graph of using the linear model (19) is a linear plane, thus it is convex. We use the parametrization (16) and apply (12) with =. In this case V can be written as V ( p I M F g)= l pj (y i+j ^y i+jji )(2) i2i i2i N + kf l ; F n k(1 ; p ) = i2i i2i N l pj (y i+j ^y i+jji )+ (1 ; p ) The optimal choice of p increases with an increase in (or equivalently an increase in ). Further, the predictor becomes linear, ie. p = 1, for large values of. To study convexity we compute the convexity measure (14) for 4 dierent values of. The measure is computed by discretizing the -set uniformly into 2121 blocks, and discretizing the v-axis into 1 equidistant intervals. The results using this approximation are shown in Table 2. In this case is a monotonically decreasing function of p.further, the computational load and the number of failures to reach the global minimum decreases with an increase in p. V 1 provides the values for the prediction error term of V using the identication data in Figs. 1 and 2. It increases quadratically with p since there is an exact match tothe identication data for p =. p V Average no. of functions calls %failto global minimum Table 2. Example 2: The eect of different p -values on the convexity measure, the computational load, and the search for the minimizing control inputs,, isgiven. The decision surfaces for dierent p values are shown in Fig. 4. As can be expected the minimizing control sequence changes with varying predictors. For large p values the minimizing control sequence is f;: :g instead of f2:1 :g as for the correct predictor.. DISCUSSION The simulation example shows that it is feasible to include convexity into the identication problem. The identication problem is aected in two ways. First, the prediction quality of the identied model decreases. This is clearly observed by viewing row V 1 in Table 2. Second, the identication problem becomes more complex. The complexity increase is, however, very modest because of the suggested parametrization (16). The example gives a merit to the proposed method since all the tested properties of the optimizationbased control problem, ie. the convexity measure,

6 u u_s Process output y time Fig. 1. Example 2: Identication input data (solid lines), validation data (broken line). Sampling time :1. the computational load, and the ability to reach the global minimum, improves with an increase in (or equivalently p ). This result can, however, not be generalized since there does exist examples where the proposed parametrization does not improve the convexity. The results on the computational load and the failure rate to reach a global minimum will depend on the optimization method used. In this work we use a Sequential Quadratic Programming method which is the basis for the Matlab routine constr (Grace 1994). The motivation for this choice is the fact that this is an extensively used mainstream method time Fig. 2. Example 2: Identication process output data. Solid and broken line corresponds to input data given by the upper and lower solid lines in the previous gure. Sampling time : CONCLUSIONS This study shows how it is possible to take into account the convexity of an on-line optimization problem when identifying a predictor. 7. REFERENCES Floudas, C. (199). Nonlinear and Mixed-Integer Optimization. Oxford University Press. Grace, A. (1994). Optimization Toolbox - User's guide. The MathWorks Inc. Johansen, T. A. (1996). Identication of nonlinear systems using empirical data and prior knowledge - an optimization approach. Automatica 32, 337{36. Lee, J. H. (1996). Recent advances in model predictive control and other related areas. In: Preprints CPC-V, Lake Tahoe, USA. Luenberger, D. G. (1969). Optimization by Vector Space Methods. JohnWiley. Mayne, D. Q. (199). Optimization in model based control. In: Preprints IFAC Symposium DYCORD, Helsingor, Denmark time Fig. 3. Example 2: Process output data. Solid and broken line correspond to process output and 1-step ahead prediction using the linear model. Control input data is given by the broken line data in Fig.3. Sampling time : \theta_p=. \theta_p= \theta_p=.2 \theta_p=.7 Fig. 4. Example 2: Decision surfaces, ie. (u u 1 ) in (18), for p = f: :2: :7g are shown. x =(2:143 1:9 1:) T.