This paper gives a brief list of commonly used direct and indirect efficient methods for the numerical solution of optimal control problems. To improve the low accuracy of the direct methods and to increase the convergence areas of the indirect methods we suggest a hybrid approach.
This paper gives a brief list of commonly used direct and indirect efficient methods for the numerical solution of optimal control problems. To improve the low accuracy of the direct methods and to increase the convergence areas of the indirect methods we suggest a hybrid approach.
This paper gives a brief list of commonly used direct and indirect efficient methods for the numerical solution of optimal control problems. To improve the low accuracy of the direct methods and to increase the convergence areas of the indirect methods we suggest a hybrid approach.
DIRECT AND INDIRECT METHODS FOR TRAJECTORY OPTIMIZATION
O. y o n S T RYK and R. BULI RS CH Mat hemat i sches Institut, Technische Universit?it Miinchen, Post f ach 20 24 20, D- 8000 Mi i nchen 2, Germany Abstract This paper gives a brief list of commonly used direct and indirect efficient methods for the numerical solution of optimal control problems. To improve the low accuracy of the direct methods and to increase the convergence areas of the indirect methods we suggest a hybrid approach. For this a special direct collocation method is presented. In a hybrid approach this direct method can be used in combination with multiple shooting. Numerical examples illustrate the direct method and the hybrid approach. Keywords: Constrained optimal control, nonlinear dynamic systems, multiple shooting, direct collocation, nonlinear optimization, hybrid approach, estimates of adjoint variables. 1. The optimal control problem Ma n y opt i mi z a t i on pr obl e ms i n aer onaut i cs and ast r onaut i cs, i n i ndus t r i al r obot i c s and i n e c o n o mi c s can be f or mul at ed as opt i mal cont r ol pr obl ems . Th e d y n a mi c s ys t em ma y be des cr i bed as a syst em of nonl i near di f f er ent i al equat i ons ~ ( t ) = f ( x ( t ) , u ( t ) , t ) , to < t < t f , (1) wi t h t 0, X(t o) and s ome Xk(tf), k e {1 . . . . . n}, gi ven and tf f i xed or f r ee. The n- ve c t or f unc t i on o f st at es x( t ) is de t e r mi ne d by an / - vect or f unct i on o f cont r ol s u(t ). Th e pr obl e m is t o f i nd f unct i ons u( t ) t hat mi ni mi z e a f unct i onal o f Ma y e r t ype J [ u ] = ~ ( x ( t f ) , t f ) wi t h O: I R n+l ---->~. ( 2) Al s o mo r e ge ne r a l t ypes o f f unct i onal s ar e possi bl e. Fo r r eal i st i c pr obl e ms it is i mpor t ant to i ncl ude pat h cons.traints. Mos t c o mmo n ar e cons t r ai nt s on t he cont r ol var i abl es C ( u ( t ) , t ) < O, to < t <- t f , ( 3) or c ons t r a i nt s on t he st at e var i abl es S ( x ( t ) , t ) <- O, to <- t < t I , J.C. Baltzer AG, Scientific Publishing Company 358 O. von St ryk, R. Bul i rsch, Me t hods f o r t raj ect ory opt i mi zat i on or mor e gener al l y S ( x ( t ) , u ( t ) , t ) < O, t o < t < t f . ( 4) In s ummar y we have m const rai nt s g ( x ( t ) , u ( t ) , t ) = ( C , S ) - ( g l . . . . . gin). For exampl e, in aeronaut i cs we oft en have position, vel oci t y and mass as state vari abl es, angl e of attack, t hrust and di rect i on of thrust as cont r ol vari abl es and we wi sh to mi ni mi ze t he t ot al fl i ght t i me t f o r to maxi mi ze t he payl oad. Typi cal const rai nt s are const rai nt s for t he thrust, dynami c pressure const rai nt s or const rai nt s on t he fl i ght pat h over cert ai n gr ound domai ns. In economi cs we may have product i on rate, amount of capital or number of empl oyed peopl e as state variables, gross i nvest ment s and expendi t ures for educat i on as cont rol vari abl es and our aim is to obt ai n full empl oyment by stable growt h. As const rai nt s we must ensur e that t he l i vi ng st andard does not fall beyond a cert ai n limit. Ther e are also appl i cat i ons to nonst andar d probl ems (cf. Fei cht i nger and Mehl mann [10]). 2. The numer i cal solution For classical probl ems and some special weakl y nonl i near l ow di mensi onal syst ems the sol ut i on can be obt ai ned anal yt i cal l y from the necessar y and suffi ci ent condi t i ons of opt i mal i t y, see e.g. sect i on 5.1. But to obt ai n a sol ut i on of dynami c syst ems descri bed by st rongl y nonl i near di fferent i al equat i ons, see e.g. sect i on 5.2, it is necessar y to use numer i cal met hods. For a fi rst sur vey t hese met hods can be cl assi fi ed into t wo types. 2.1. NECESSARY CONDITIONS FROM CALCULUS OF VARIATIONS The i n d i r e c t met hods are based on t he cal cul us of vari at i ons or t he ma xi mum pri nci pl e. Under cert ai n assumpt i ons (see e.g. Br yson and Ho [1], Hest enes [14]) t he f ol l owi ng first or der necessar y condi t i ons for an opt i mal t raj ect ory are valid: Ther e exi st an n- vect or funct i on of adj oi nt vari abl es ~(t) and an m- vect or f unct i on v ( t ) , such that with the so-cal l ed Hami l t oni an funct i on H = ~Tf + vTg (5) a mul t i - poi nt boundar y val ue probl em in canoni cal form with pi ecewi se def i ned di fferent i al equat i ons results f or t o < t < tf: ~H .~ = ~ = f , ( 6) O. von Stryk, R. Butirsch, Met hods f o r trajectory optimization 359 ~, = ~H _ ~ T ~ f _ v T~)__g.g ---~-X = -~X Ox' (7) o _ g. ( 8 ) The opt i mal cont rol is det ermi ned by mi ni mi zi ng H with respect to u. For exampl e, for H nonl i near in u we may use for to < t < t f 0 H= ~ T b f + v T Og 3u ~ ~ = 0 (9) and t he Legendr e- Cl ebs ch condi t i on ~ZH/~u,~3uk positive semidefinite. We have addi t i onal (trivial) differential equations for t) and the swi t chi ng points qi in whi ch the const rai nt s become active or inactive. Additional constraints at initial, end and interior poi nt s may hold. The numeri cal procedure described in sections 3 and 4 is also able to handl e opt i mal singular arcs. However, for the sake of simplicity, this case is not di scussed here. To obt ai n sol ut i ons from these necessary conditions we may use met hods whi ch are based on t he special structure of these necessary conditions, e.g. so-called gradi ent met hods (see e.g. Gottlieb [ t 2], Tol l e [25], Bryson and Ho [ 1 ], Chernousko and Lyubushi n [8] and Miele [18]). Alternatively, we obtain the controls u( t ) from 3 H/ ~ u = 0 analytically or numeri cal l y usi ng Newt on' s met hod, and we may use a met hod for the solution of general boundary value probl ems such as the multiple shoot i ng met hod (see e.g. [2,23]) or a col l ocat i on met hod (see e.g. Di ckmanns and Well [9]). Cont rary to ot her met hods the multiple shooting met hod has the advantage that all ki nds of constraints are allowed and very accurate results can be obtained. However, a rather good initial approxi mat i on of the optimal trajectory is needed and a rat her large amount of work has to be done by the user to deri ve the adjoint differential equations. Moreover, the user has to know a priori the switching structure of t he constraints. This can be derived by means of homot opy techniques. For a description of useful techniques in conjunction with multiple shooting for the numerical sol ut i on of const rai ned opt i mal control problems the reader is referred to [3]. All in all, the user must have a deep insight into the physical and mathematical nat ure of the opt i mi zat i on problem. 2.2. DIRECT SOLUTION OF THE OPTIMAL CONTROL PROBLEM In d i r e c t approaches the optimal control problem is transformed into a nonlinear pr ogr ammi ng probl em. In a first approach, this can be done with a so-called direct shoot i ng met hod t hrough a paramet eri zat i on of the controls u( t ) . For this we choose u ( t ) from a finite di mensi onal space of control functions and use explicit numeri cal integration to satisfy the differential eqs. (1), see e.g. Wi l l i amson [27], Kraft [17], Horn [15], Bock and Plitt [5], to cite only a few of the many papers. 3 6 0 O. yon St ryk, R. Bul i rsch, Me t hods f o r t raj ect ory opt i mi zat i on In a second approach, u(t) and x(t) are chosen from finite di mensi onal spaces u ~ span {/~1 . . . . . /~p }, U = ~ O~i~t i , O~ i E I R, ( 1 0 ) i=1 q x ~ span{~l . . . . . 3~q}, X ----- Z ~ j . ~ j , ~ j ~. I R. ( 1 1 ) j = l Compar ed wi t h t he first appr oach this has t he addi t i onal advant age that t he comput at i onal l y expensi ve numeri cal i nt egrat i on of the differential eqs. (1) can be avoided. In common approaches pi ecewi se pol ynomi al approxi mat i ons are used (cf. Renes [22], Kraft [17], Hargraves and Paris [13]). Approxi mat i ons as finite sums of the Chebyshev expansi ons of x ( t ) and u ( t ) are also possi bl e but not easy to handl e in the presence of path constraints (see e.g. Vl assenbroek and van Dooren [26]). The differential eqs. (1) and the path constraints (4) are onl y satisfied at discrete points. The resulting nonl i near program is Mi ni mi ze ~ ( I 0 , Y = ( a l . . . . . ap, f l ~ . . . . . ~q, t f ) , y ~ I R p +q+l, (12) subject to a(Y) = 0, b ( Y ) < O. (13) The differential equations, initial and end poi nt condi t i ons and pat h const rai nt s ent er the const rai nt funct i ons a and b of t he nonl i near program. The nonl i near pr ogr ammi ng probl em is sol ved by usi ng ei t her a penal t y funct i on met hod or met hods of augment ed or modi fi ed Lagrangi an funct i ons such as sequential quadratic pr ogr ammi ng met hods. The advant age of the direct approach is that the user does not have to be concerned with adj oi nt variables or swi t chi ng structures. One disadvantage of direct met hods is that they produce less accurate solutions t han indirect met hods: By sol vi ng numeri cal l y several di ffi cul t opt i mal cont rol probl ems from aeronautics, we found that in practice t he mi ni mum funct i onal val ue is obt ai ned with relative l ow accuracy (i.e. errors of about one percent). Increasi ng the di mensi on of the finite di mensi onal space does not necessari l y yield bet t er values for the ext remel y compl i cat ed probl ems arising from aerodynamics. However, this "quantity" of one percent can be a crucial part of the payl oad in a space flight mi ssi on (cf. Callies et al. [6], Chudej et al. [7]). A second di sadvant age is that the discretized opt i mal control probl ems have somet i mes several mi ni ma. Appl yi ng the direct met hods to t he discretized probl em t hey oft en end up in one of these "pseudomi ni ma". Thi s sol ut i on, however, can be quite a step away from the true sol ut i on satisfying all the necessary condi t i ons from variational calculus resulting, e.g., in a 20 percent worse funct i onal value. Exampl es of such probl ems are reported by Bock et al. [16] and by [24]. To overcome these disadvantages it is necessary to combi ne direct with indirect met hods. In the following, we present a special direct met hod used in conj unct i on with mul t i pl e shoot i ng. O. v o n S t r y k , R . B u l i r s c h , M e t h o d s f o r t r a j e c t o r y o p t i m i z a t i o n 361 3. A di rect me t hod The basi c ideas of this di rect col l ocat i on met hod wer e out l i ned by Renes [22], Kraft [17], Har gr aves and Pari s [13]. Several additional practical feat ures are added such as sel ect i on of initial values, node selection, node refi nement , accur acy check and conver gence i mpr ovi ng feat ures as, for exampl e, scaling of variables, additional const rai nt s, speci al t r eat ment of angles [24]. A di scret i zat i on of t he t i me interval t o = t I < t 2 < . . . < t N = t f (14) is chosen. The par amet er s Y of the nonl i near program are the val ues of cont rol and state vari abl es at the grid points t j , j = 1 . . . . . N, and t he final time tf: Y = (u(t l) . . . . . u ( t N ) , X ( t 1 ) . . . . . X ( t N ) , t N) E I R N(t + n)+ I (15) The cont rol s are chosen as pi ecewi se l i near interpolating funct i ons bet ween u ( t j ) and u ( t j + 1) for t j <_ t < 5 + I: t - - t j Uap p ( t ) = U ( t j ) + tJ+l _ t------~ ( U ( t j + l ) - U ( t j )). (16) Z61 o 0 ~j +! Fig. 1. Approximation of u(t). The states are chosen as cont i nuousl y di fferent i abl e and pi ecewi se cubi c funct i ons b e t w e e n x ( t j ) and x ( t j + l ) , ~app(S):= f ( x ( s ) , u ( s ) , s ) , s = t j , t j + l , for t j <_ t < t j + l , j = l . . . . . N- l , ~ , . j t - t j ~ k c 6 = x ( t j ) , c = h i D , (17) (18) (19) 362 O. v o n St r y k , R. Bu l i r s c h , Me t h o d s f o r t r a j e c t o r y o p t i mi z a t i o n X - , . , 0 0 Fig. 2. Approximation of x(t) and implicit integration scheme. ci2 = - 3 x ( t j ) - 2 h j 3 ~ + 3 x ( t j + l ) - h j ~ + l , ( 20) = 2 x ( t j ) + h s f j - x ( t i + l ) + h s f ( 2 1 ) wher e f j : = f ( x ( t j ) , u ( t j ) , t j ) , h j : = t j + 1 - t j . The states are pi ecewi se cubi c Her mi t e i nt erpol at i ng funct i ons. Reader s who are not f ami l i ar wi t h this t erm are referred to c ommon t ext books of Numer i cal Anal ysi s, such as e.g. [23]. The const rai nt s in t he nonl i near pr ogr ammi ng pr obl em are ,, t he path const rai nt s of t he opt i mal cont rol pr obl em at t he grid poi nt s tj: g ( x ( t j ) , u ( t j ) , t j ) < O, j = 1 . . . . . N , * the speci fi ed val ues of the state vari abl es at initial and t ermi nal t i me, * and f ( X a p p ( t ) , U a p p ( t ) , t ) - J q p p ( t ) = 0 for t = t j + tj+ _.________~l j = 1, N- 1 . ( 22) , **. ~ Fur t her on, t he i ndex "app" f or approxi mat i on will be suppressed. Thi s way of di scret i zi ng x(t) has the advant age that not onl y t he number of par amet er s of the nonl i near pr ogr am is reduced (because :~(t/), j = 1 . . . . . N, are not par amet er s) but also t he number of const rai nt s is reduced by this i mpl i ci t i nt egrat i on scheme (because t he const r ai nt s i c ( t j ) = f ( x ( t j ) , u ( t j ) , t j ) , j = 1 . . . . . N , are f ul f i l l ed by t he chosen approxi mat i on). Ot her ways of di scret i zi ng x ( t ) do not have this propert y. For t he t i me bei ng, t he nonl i near program is sol ved by usi ng a met hod based on sequent i al quadr at i c pr ogr ammi ng due to Gill et al. [11]. Thi s di rect col l ocat i on met hod has been successf ul l y applied to di ffi cul t const rai ned opt i mal cont rol pr obl ems such as t he mi ni mum accumul at ed heat descent t raj ect ory of an Apol l o capsul e wi t h hei ght const rai nt [24]. Conver gence was achi eved even in the case t hat no i nf or mat i on was gi ven about t he opt i mal t raj ect ory. Onl y i nf or mat i on of t he gi ven dat a was used. O . v o n S t r y k ~ R . B u l i r s c h , M e t h o d s f o r t r a j e c t o r y o p t i m i z a t i o n 363 Approximations of the minimum functional value and of the state variables with a relative error of one percent were achieved and were quite satisfactory. Approximations of the optimal control variables, however, were worse compared to the "exact" solution (see figs. 4 and 8). Also the disadvantages of direct methods described above have been observed. 4. The hybrid approach To overcome the disadvantages of the direct method it would be desirable to combine the good convergence prope~ies of the direct method with the reliability and accuracy of the multiple shooting method. But the switch between both methods is not easy as it is necessary to set up the adjoint differential eqs. (7) and the optimal control laws (9). Also a proper choice of the multiple shooting nodes, initial values of the adjoint variables ~,(t) and the variables v ( t ) and the switching structure is needed in advance. As for the combination of the methods, the grid points of the direct method yield a good choice for the positions of the multiple shooting nodes. For example, in fig. 6 the final distribution of the 25 grid points obtained by the node selection and refinement described in [24] is marked by crosses on the time axis. As an additional advantage, reliable estimates for the adjoint variables (that do not explicitly appear in the direct formulation) can be obtained from the parameters and the Lagrangian function of the nonlinear program. We assume sufficient differentiability for all functions and (only for the purpose of illustration) n = 1, l = 1, m = 0. Evaluation of the approximations of u ( t ) and x ( t ) of the direct method from section 3 at the center t i + 1/2 : = ( t l + t i + 1)/2 of each discretization interval yields U ( t i + l l 2 ) = 1 ( Ui + Ui+l ) , ( 2 3 ) 1 ( X i + Xi + l ) + t i + 1 8 t i ( f ( x i , u i , t i ) - - f ( X i + l , U i + l t i + l ) ) , ( 2 4 ) X ( t i + l t 2 ) = ~ _ 3 ( X i + l - - X i ) 1 ( f ( x i , u i t i ) + f ( X i + l , U i + l , t i + l ) ) ' ( 2 5 ) " ~ ( t i + l / 2 ) - - 2 ( t i +l -- t i ) - - -~ i =1 ..... N- I . Here the notation u i := U ( t i ) , X i := X ( t i ) , i = 1 . . . . . N , (26) has been used, where u~ and x i are the parameters of the nonlinear program. Furthermore the relations ~ u ( t i + l / 2 ) - 0 = ~ u ( t i - 1 / 2 ) (27) ~ x i ~ x i ' 364 O. von Stryk, R. Bulirsch, Methods f or trajectory optimization O X (ti +l l 2) = 1_ _ + ti +l - t i Of ( x i , u i , t l ) (28) ~x i 2 8 bx ' ~ X ( t i - l / 2 ) = 1 _ _ _ t i - - t i - I ~ f ( x i , u i , t i ) ( 29) Oxi 2 8 Ox " O:c(ti+l/2) 3 1 Of ( x i , u i , t i ) = ( 3 0 ) Ox i 2(ti+ 1 - t i ) 4 Ox ' ~ X ( t i - l l 2 ) 3 1 Of ( x i , u i , t l ) = (31) Ox i 2 ( t i - t i _ l ) 4 Ox are needed. The Lagr angi an of t he associ at ed nonl i near pr ogr am from sect i on 3 is N-1 L ( Y , # ) = q a ( XN, t N) - ~_~#j ( f ( x( t j +l / 2) , u( t j +l / 2) t j +l / 2) - - 2( t i +l / 2) ) , ( 32) j=1 wi t h # = (/.q . . . . . ]AN- I) IRN - I A sol ut i on of the nonl i near pr ogr am fulfills the necessar y first or der condi t i ons of Kuhn and Tucker . In detail, we fi nd among ot hers f or i = 2 . . . . . N- 1 o = ) --].ti(~f(x(ti+l/2)'U(ti+ll2)'ti+ll2) ~JT(fi+I/2 ) 1 3 x i ~ x i " (33) Subst i t ut i on of ( 2 7 ) - ( 3 1 ) into (33), the chai n rule of di fferent i at i on, and s ome basi c cal cul at i ons l ead to ~L 3 ( #i I -ti-1 1 1 ~ f ( x i , u i , t i ) ~ X i = t i + 1 - - t i t i -- ti_ 1 4 ~X (#i-1 + # i ) 1 ( ~f ( x ( t i - l l 2) ' U( t i - l / 2) , t i - l / 2) Of (x(l i +l l 2)' U(t i +l l 2)' t i +l / 2) 1 - 7 #i -1 bx ~ #i bx O f ( x i , u i , t i ) I t i - - t i - 1 t i + t i 1 4 ~x - - - - i f - - #i-1 ~ - #i . (34) For conveni ence, we now suppose an equi di st ant grid, i.e. ti+l - ti = h = tf - to N- 1 ' i = l . . . . . N- 1 . (35) O. v o n S t r y k , R. Bu l i r s c h , Me t h o d s f o r t r a j e c t o r y o p t i mi z a t i o n 365 Int roduci ng the abbreviations J~ : = f ( x i ' u i ' t i ) , J ~ + l / 2 : = f ( x ( t i + l l 2 ) , U ( t i + i l 2 ) , t i + l l 2 ) , (36) eq. (34) can be rewri t t en as OL = - 3 ( # i - #i-1) Oxi 43x (/zi-1 q - # i ) - - ' 2 ~ i - l ~ - " - f f ~ x - - ~ i ~X i " -ff-~- t#i-1 - #i ). (37) By usi ng Tayl or ' s t heorem / - + o ( h ) , / 2 - - xOZ + O ( h ) ' (38) we obt ai n an adj oi nt difference equation for #;, valid for i = 2 . . . . . N- 1, OL _ 3 # i - # i - 1 Oxi 2 h 3 (#i + # i - 1 ) - ~ + O ( h ) 4 = 0. (39) By keepi ng t = t i fixed and letting h ---> 0 by increasing the number of grid points N---> ,,~, eq. (39) converges to the adjoint differential equation 12 ( t i ) = - I d ( t i ) O f ( x ( t l ) ' u ( t i ) ' t i ) Ox So we use (40) ~ t , ( t i + l l 2 ) = O ' . ( - - # i ) , cr =cons t . >0 , i = 2 . . . . . N- l , (41) as an est i mat e for ~(t) with a scaling factor or. For all optimal control probl ems of Mayer type, tr can be easily det ermi ned by using the additional end poi nt condi t i on for 2(ty) from cal cul us of variations. Relation (41) was used in the exampl es of sect i on 5. Obvi ousl y, an estimate for the mul t i pl i er funct i on v ( t ) appearing in ( 5) - ( 9) in the presence of constraints g can also be obtained by this approach. The exampl es show that the switching structure for the constraints on the state variables will be approxi mat ed in a satisfactory way by a sol ut i on of t he direct met hod. However, the swi t chi ng structure of the constraints on the cont rol variables will not be approxi mat ed as well due to the bad approxi mat i on of t he control variables. At this point, further refinements are required. Nevert hel ess, this hybrid approach has been successfully applied to several test exampl es and new real-life problems with unknown solutions as, for exampl e, the maxi mum range trajectory of a hanggl i der [4]. 366 O. von Stryk, R. Bu l i r s c h , Me t h o d s f o r t r a j e c t o r y o p t i mi z a t i o n 5. Numeri cal exampl es 5.1. THE BRACHISTOCHRONE PROBLEM As a first exampl e and in order to illustrate the propert i es of t he di rect met hod for est i mat i ng adj oi nt variables the wel l -known classical Brachi st ochrone probl em was chosen. Here, the numeri cal sol ut i on from the direct col l ocat i on met hod is compar ed wi t h the exact solution. As for the formul at i on of the probl em the not at i on of Bryson and Ho [1] is used, see sect i on 2.7, probl em 6 (the gravity acts in t he di rect i on of the y-axis): Mi ni mi ze J[O] = t f , subj ect to the differential equations (42) Jc( t ) = ~ V ~ y ( t ) cos~9(t), g = 9.81, (43) ~(t) = ~y(t) sin O(t), 0 <t <t f , and the prescribed values at initial and terminal t i me (44) x ( O ) = O, x ( t f ) = 1, (45) y( 0) = 0, y ( t f ) free. (46) The cont rol O(t) is the angle of the t angent of t he trajectory ( x ( t ) , y ( t ) ) . The necessary condi t i ons, derived from calculus of variations, can be sol ved analytically; and we obtain for the state and control variables of the opt i mal trajectory in 0 < t <_ tf: 1 ( cot - sin cot), c o=~f - ~, (47) x ( t ) = - y y ( t ) = 2 s i n 2 ( 2 t ) , (48) 1 O( t ) = ~ ( ~ - cot), (49) and the adj oi nt variables ~,x ( t ) = - ~ ~gg , ;I.y (t ) = ~,x COt ( 2 t ) The final t i me and mi ni mum funct i onal value is t f = ~ g g . (50) (51) (52) O. von Stryk, R. Bulirsch, Met hods f o r trajectory optimization 367 -y OOoi - . 2 5 -. 5 0 - . 7 5 -I .00 27 ~ "'. .25 ,50 .75 l . O 0 Fig. 3. The (x(t ), - y(t)) trajectory in the plane. 1 . 5 0 1, 25 1. 00 . 7 5 . 5 0 . 2 5 0. 00 0. 00 l o . 20 . 30 . 40 . 50 Fi g. 4. Th e o pt i ma l conta'ol O ve r s us t i me. 368 O. von Stryk, R. Bulirsch, Methods f or trajectory optimization o o o i - 1. 00 - 2. O0 - 3. 00 - 4 . O0 L oo . t i i Fig. 5. The adjoint variable ~ versus time. Fi gures 3 to 5 show the initial trajectory (- - -) of the direct met hod for 5 equi di st ant grid points, the sol ut i on of the di rect met hod ( . . . . . ) for 10 grid poi nt s and t he analytic sol ut i on (-). In the figures there is no visible di fference bet ween the approxi mat ed and the exact state variables x and y. The mi ni mum funct i onal val ue is achi eved wi t h an error of 7/1000 percent. Obviously, the si ngul ari t y of A.y(t) at t = 0 affects t he quality of the approxi mat i on of O(t) in t he nei ghbor hood of t = 0. It shoul d be not ed that due to the singularity of ~( t ) at t = 0, the probl em cannot be sol ved by mul t i pl e shoot i ng when usi ng this formul at i on. An addi t i onal r ef i nement as, e.g., a local finite Tayl or series is t herefore necessary. 5.2. THE APOLLO REENTRY PROBLEM To illustrate the propert i es of t he hybri d approach we sol ve anot her well- known but rather difficult probl em of fi ndi ng the mi ni mum accumul at ed heat descent trajectory of an Apol l o capsule (cf. [23,24] and Pesch [21]). The di fferent i al equat i ons of this t wo di mensi onal model are for 0 < t < t/: ~, _ S go sin y 2 m P ( ~ ) V 2 C D ( U ) (1+ ~) 2, (53) = Rs i n y, (54) = l - - ~ c o s y , (55) O. yon Stryk, R. Bulirsch, Met hods f o r trajectory optimization 369 ~" = - ~ m p ( ~ ) v c t ` ( u ) + d t = C V 3 ~ P ~ ) , vcos 7 go cos 7 (56) R( I + ~ ) v ( l + ~) 2' (57) where the abbreviations c o ( u ) = Coo + Cot ` cos u, Coo = 0. 88, Cot ` = 0. 52, (58) c L ( u ) = ct` o si nu, ct` o = - 0. 505, (59) P ( ~ ) = P o e x p ( - f l R ~ ) , Po = 2. 3769x10- 3, (60) S = 50000, go = 3.2172 1 0 - 4 , R = 209.0352, (61) r n ~ = 1 0. 235' c = 2 0 , N= 4 , (62) have been used. The state variables are the velocity v, t he normalized hei ght ~, the range over ground 5, the flight path angle 7and the accumul at ed heat q. The control variable is the angle of attack u. The final time tf is free. The prescribed values at initial and t ermi nal t i me are v ( o) = 0. 35, ~ ( o ) = 4 / R , ~(o) =o, 7( 0) = - 5. 75. ~/180, q(O) = o, v ( t / ) =0 . 0 1 6 5 , ~ ( t f ) = 0. 75530/ R, i f ( t / ) = 51.6912, 7( i f ) free, q ( t : ) free. The funct i onal to be mi ni mi zed is J[ u] = q( t / ) . The adj oi nt differential eqs. (7) from calculus of variations are ; t , v = - 3 c v 2 ~ ' - ' - ~ - ' A , q f p ( ~ ) ' ~ + 2 - - ' ~ p ( ~ ) ( 2 v / t , v c o ( u ) - / t , r c t , ( u ) ) ~ s i n T~ cos r (~'~' + go z ~ + , ~ ) - - - T - , ~ , - T ~ ( t , - - k - - v 2 ( 1 + , ) (63) (64) (65) 3 7 0 O. yon Stryk, R. Bulirsch, Methods for trajectory optimization ~ _ 1- ~ 3 f f ~- ~ _ 2~R p( ~) ( V 2A vC o( U ) _ - -~pK cv ~l T , ~q vZr cL (u)) o in, cos,' 'o - 2 / ~ v ( 1 + ~ ) 3 ( 1 + ~ ) 2 7 ( 1 + ~ - - - - - - ~ ' ;/,;. = o , ) ~ 7 = 1 - - ~ R ~ r v ( l + ) ~(1+)2 ~ ; L , , ~q = 0 . Furthermore, we have the additional conditions /~r (tf ) = O, /~q(tf ) = 1, H ( tf ) = O, and the optimal control law = ( bcL (u) ~H S p( ~) v )],~, bu 2m bu Oco u) v a ,,, ~ u ) = 0 . (66) (67) (68) (69) (70) ( 7 0 150. O0 I00.00 50. O0 O. O0 - 0.00 100 O0 200. O0 300. OO 400. O 0 Fig. 6. The hei ght o f the caps ul e versus ti me. t a , O. von Stryk, R. Bulirsch, Methods f or trajectory optimization 371 . I O E - O 1 . 5 0 E - 0 2 O . O O E + O 0 0 - . 5 0 E - 0 2 - . I O E - O 1 - . 1 5 E - O 1 - . 2 0 E - O 1 ! )0 / 1 0 0 . 0 0 / / / / / / / / i / "%-- " , . , . ~. , 2 0 0 . O0 . , ~ . ~ ' ~ ' ~ ' ~ ' - 3 0 0 . O0 ~ " . , , , . , , 0 0 . O0 l ~ l \ Fig. 7. The adjoint variable of the height versus time. [oZeg] 1 8 0 . 0 0 9 0 . 0 0 0 . 0 0 0 - 9 0 . 0 0 - 1 8 0 . 0 0 i.~.~ .~ i i i i i I , I I I 0 0 200. O 0 800. O 0 i400. O0 il A i I \ ~ I ! % / \ / \ / . . . . . . . % .,,- ",,. . / Fig. 8. The angle of attack versus time. 372 O. yon Stryk, R. Bulirsch, Methods f or trajectory optimization An initial trajectory for the iteration process is generat ed at ni ne equi di st ant grid poi nt s from the gi ven prescribed values of t he state variables at initial and final time. The direct col l ocat i on met hod converges in several refi nement steps. Addi t i onal gri d poi nt s had to be added in order to obtain a 25 grid poi nt solution. By an aut omat i c procedure the number and posi t i ons of mul t i pl e shoot i ng nodes and the val ues of state and adj oi nt variables are obt ai ned from the sol ut i on of t he di rect met hod. From t hese starting values the mul t i pl e shoot i ng met hod (cf. [2, 19, 20, 23] ) converges to a very accurate solution. The error in the achi eved funct i onal value and t he final t i me of the sol ut i on of t he di rect met hod compar ed with t he mul t i pl e shoot i ng sol ut i on was about one percent . Fi gures 6 to 8 show the hei ght of the capsule, the adj oi nt variable of t he hei ght and the cont rol angle of attack, t he initial (- - -) trajectory, the result of t he di rect met hod ( . . . . . ) and the sol ut i on of the mul t i pl e shoot i ng met hod (-). In addition, fig. 6 also shows t he initial and final di st ri but i on of the grid poi nt s of t he direct met hod marked by crosses. 6. Conclusions We have demonst rat ed that a combi nat i on of di rect and indirect met hods is a very pr omi si ng way to obtain the numeri cal sol ut i on of nonl i near opt i mal cont rol probl ems. The swi t ch bet ween bot h met hods in the case of general nonl i near const rai nt s on the cont rol and state variables is current l y studied. First numeri cal results are encouragi ng. Nevert hel ess, t he sol ut i on of a di ffi cul t real-life opt i mal cont rol probl em cannot be obt ai ned wi t hout any i nsi ght into the mat hemat i cal and physi cal nat ure of the sol ut i on of the opt i mal control probl em. References [1] A.E. Br yson and Y.--C. Ho, Applied Optimal Control, Rev. Printing (Hemisphere, New York, 1975). [2] R. Bulirsch, Die Mehrzi el met hode zur numeri schen LOsung yon nichtlinearen Randwert probl emen trod Auf gaben der optirnalen Steuerung, DLR, Oberpfaffenhofen, Ger many, Repor t of the Carl- Cranz-Gesel l schaft (1971). Reprint: Depart ment of Mat hemat i cs, Muni ch Uni versi t y o f Technol ogy (1985). [3] R. Bulirsch, F. Montrone and H.J. Pesch, Abort landing in the pr esence of a windshear as a mi ni max opt i mal control probl em. Part 1: Necessary conditions. Part 2: Mul t i pl e shoot i ng and homot opy, Schwer punkt pr ogr amm der DFG: Anwendungsbezogene Opt i mi erung und Steuerung, Repor t No. 210 (1990), ]. Opt i m. Theor y Appl. (1991), to appear. [4] R. Bulirsch, E. Nerz, H.J. Pesch and O. yon Stryk, Maxi mum range trajectory of a hanggl i der through a thermal, in preparation. [5] H. G. Bock and K.J. Plitt, A multiple shooting algorithm for direct solution of opt i mal control pr obl ems, IFAC 9th World Congress, Budapest, Hungary (1984). O. von Stryk, R. Bulirsch, Methods f or trajectory optimization 373 [6] R. Callies and R. Bulirsch, Optimale Flugbahn einer Raumsonde zum Planetoiden Vesta, Jahrbuch der DGLR 11(1990)895-904. [7] K. Chudej, R. Bulirsch and K.D. Reinsch, Opt i mal ascent and staging of a two-stage space vehicle syst em, Jahrbuch der DGLR I ( 1990) 243- 249. [8] F.L. Chernousko and A.A. Lyubushin, Met hod of successive approximations for opt i mal control pr obl ems (survey paper), Opt. Contr. Appl. Meth. 3( 1982) 101- 114. [9] E,D. Di ckmanns and K. H. Well, Approximate Solution of Optimal Control Problems Using Third Order Hermite Polynomial Functions, Lecture Notes in Computational Science, Vol. 27 (Springer, Heidelberg, 1975) pp. 158- 166. [10] G. Feichtinger and A. Mehlmann, Planning the unusual: Applications of control theory to nonstandard pr obl ems, Act a Appl. Math. 7( 1986) 79- 102. [11] P.E. Gill, W. Murray, M.A. Saunders and M.H. Wright, Us er ' s Guide for NPSOL (version 4.0), Repor t SOL 86-2, Department of Operations Research, Stanford University, California (1986). [12] R.G. Gottlieb, Rapi d convergence to opt i mum solutions using a min-H strategy, AIAA J. 5(1967) 3 2 2 - 3 2 9 . [13] C. R. Hargraves and S.W. Paris, Direct trajectory optimization using nonlinear programmi ng and collocation, AI AA J. Guidance 10(1987)338-342. [14] M. R. Hestenes, Calculus of Variations and Optimal Control Theory (Wiley, 1966). [15] M. K. Horn, Solution of the opt i mal control probl em using the software package STOMP, 8th IFAC Workshop on Control Applications of Nonlinear Programming and Optimization, Paris (1989). [16] J. Konzel mann, H. G. Bock and R.W. Longrnan, Ti me optimal trajectories of elbow robots by direct methods, Proc. AIAA Guidance, Navigation, and Control Conf., Boston (1989) AIAA Paper 89-3530-CP. [17] D. Kraft , On converting optimal control probl ems into nonlinear programmi ng codes, in: NATO ASI Series, Vol. F15, Computational Mathematical Programming, ed. K. Schittkowski (Springer, 1985) pp. 2 6 1 - 2 8 0 . [18] A. Miele, Gradi ent algorithms for the optimization of dynamic systems, in: Control and Dynamic Systems, ed. C. T. Leondes, Vol. 16 (Academic Press, 1980) pp. 1 - 5 2 . [19] H. J. Oberl e, Numeri sche Berechnung optimaler Steuerungen yon Heizung und Kt~hlung ftlr ein realistisches Sormenhausmodell, Habilitationsschrift, Munich University of Technol ogy, Ger many (1982), Report No. M8310 (1983). [20] H.J. Ober l e and W. Gr i mm, BNDSCO - a program for the numerical solution of optimal control problems, Deutsche Forschungs- und Versuchsanstalt ftir Luft- und Raumfahrt (DLR), Oberpfaffenhofen, Ger many, DLR IB 515-89/22 (1990). [21] H.J. Pesch, Real -t i me comput at i on of feedback controls for constrained opt i mal control, Part 2: A correction met hod based on multiple shooting, Optim. Contr. Appl. Meth. 2( 1989) 147- 171; see also: Schwerpunkt programm der DFG: Anwendungsbezogene Optimierung und Steuerung, Reports no. 8 and 14 (1987). [22] J.J. Renes, On the use of splines and collocation in a trajectory optimization algorithm based on mat hemat i cal programmi ng, National Aerospace Lab. Amsterdam, The Netherlands, NLR-TR78016 U (Sept. 1978). f23] J. St oer and R. Bulirsch, Introduction to Numerical Analysis, 2rid rev. ed. (Springer, 1983). i24] O. von St ryk, Ein di rekt es Verfahren zur n~iherungsweisen Bahnopt i mi erung yon Luft - und Raumfahrzeugen unter Berticksichtigung yon BeschrL, akungen, (a) Di pl oma Thesis, Department of Mathematics, Munich University of Technology, Germany (May 1989); (b) Lect ure at the Institute for Flight Systems Dynamics, DLR, Oberpfafferthofen, Ger many (June 16th 1989); (c) Annual Scientific Conf. of the GAMM, Hannover, Germany (1990), Zamm 71, 6( 1991) T705- T706. [25] H. Tol l e, Optimierungsverfahren (Springer, Berlin, 1971). [26] J. Vl assenbroek and R. van Dooren, A Chebyshev technique for solving nonlinear opt i mal control probl ems, I EEE Trans. Aut. Contr. AC-33(1988)333-340. [27] W. E. Wi l l i amson, Use of pol ynomi al approximations to calculate suboptimal controls, AI AA J. 9(1971)2271 - 2273.