Professional Documents
Culture Documents
7
7
7
250. Columns 8 and 9 of Table 4 present our proposed solution algorithm's average and maximum computational times. We explain the results shown in the last to columns of Table 4 in the next subsection Columns 2 and 3 of Table 5 show the average and max- imum percentage difference of the objective value of the solution returned by LINGO’s local solver from the prob- lem’s global optimal solution objective value. The next two columns of the same table show the average (rounded to the nearest integer) and the maximum number of Feasible value combinations of variables z_ and zy. OF course, these results are always the same for both the solution algorithm that we propose and the decoupling LINGO model. We explain the results shown in the last two columns of Tables 5 in the next subsection The superiority of the solution algorithm that we propose ‘becomes immediately clear, because its computational times, are significantly lower than those of both LINGO models. As the results of Table 4 demonstrate, the computational savings increase considerably for large scale problem instances, for Which the application of LINGO appears impracticable. This [Naval Research Logistics DOT 10.1002/nav is partially due to the fact that the increase in the number of feasible value combinations of variables zy and zy is quite ‘moderate as the problem size increases, and partially due 10 the eficiency of Procedure Sweep. The variability ofthe solu- tion times appears higher in the case of LINGO than in the ccase ofthe algorithm that we propose. Additionally, the o inal LINGO model with the local solver invoked has a rather ‘unusual behavior, because its average computational require- ments for N= 250 are lower than those for N = 150 and N = 200, This is an indication thatthe total computational effort also depends on the specific characteristics of each problem instance besides its size. With the exception of the problems with = 50, the average computational require- meats of the decoupling LINGO model are higher than those of the original LINGO model with the global solver invoked. ‘This is an indication that the significant computational sav- ings of our solution algorithm should be attributed more to the efficiency of Procedure Sweep than to the decoupling of the original problem into smaller subproblems alone. Asexpected, the computational requirements of LINGO's local solver are significantly lower than those of the global solver. This comes at a price, however, as columns 2 and 3 of Table 5 verify, because the objective value of the solution retumed by the local solver is on the average approximately between 100 and 550% higher than that of the global opti mal solution. In the worst case, this percentage difference increases to ~2000%. One way to improve the quality of the solutions retumed by LINGO’ local solver isto increase the ‘number of “multstart solver attempts” on the “Global Solver” tab, but thisalso increases the computational time required for termination. In general, as this number increases, the results, retumed by the local solver resemble those retumed by the. ‘global solver in terms of solution quality and! computational requirements 5.3. Algorithmic Enhancements ur extensive computational experience with the proposed solution algorithm and LINGO has provided considerableKozanidis, Gavranis, and Kostaelou: Flight and Maintenance Planning of Mission Aircraft 21 evidence suggesting that the two-dimensional “cost-matrix” ‘with rows the feasible values of variable z,, columns the fea- sible values of variable zp. and elements the optimal total deviation index values for each particular combination of sq and zg, may possess a special type of convexity called row and column convexity, More specifically, let TC(g.) be the problem's optimal total deviation index value for the combination wit g and z, = a. The corresponding matrix is row convex if TC(g,a) < TC(g,a + 1) implies TC(g,a +1) = TC(g.a +1) for every feasible / > 2 and if TC(g,a) < TC(g,a—1) implies TC(g.a—1) < TC(g.a—i) for every feasible / > 2. Similarly, the corresponding table is column convex if 7C(g.a) < TC(g + l.a) implies TC(g + la) = TC(x +/,a) for every feasible / > 2 and if TC(g.a) < TC(g—l,a) implies TC(g—1.a) = TC(g—i.a) For every feasible i > 2. Despite numerous and tedious attempts, we have not been able to develop a formal mathematical proof tI ishes the validity of this property. On the other hand, despite exten- sive experimentation, we have not been able to discover a single counterexample that disprovesit either. IFthis property is indeed valid, then we can exploit it to improve the com- putational performance of the proposed solution algorithm considerably. To show this, we developed a simple modifica- tion of this algorithm, which does not compute the optimal total deviation index value for every feasible combination of q and zy; instead, this algorithm computes this value for the middle element of each row (or column using a specific sim- plisticrute) of the cost-matrix and terminates its search within the Same row (or column) as soon ast has established, assum- 1 that the cost-matrix is indeed row and column convex, that ‘no further improvement on the objective ean be accomplished. inthe same row (or column). After this procedure is repeated for all rows (or columns) of the matrix, the best incumbent solution is returned by the algorithm. The last wo columns, of Table 4 present the average and maximum computational times of this modified algorithm, and the last two columns of ‘Table 5 present the average and maximum number of value ‘combinations of variables z, and z, for which this algorithm computed the total deviation index. These results show that if the cost-matrix is indeed row ‘and column convex, then the solution algorithm that we pro- pose can be significantly expedited by cleverly incorporating this property into its original design. The modified algo- rithm returned the global optimal solution in every problem instance out ofthe 200 on which it was applied, giving us. for ‘one thing, strong evidence that this is indeed true, Given this, intuition, we believe that future research should investigate ‘whether this property is actually valid or not, Ifit turns out 10 be, asuitable modification ofthe proposed solution algorithm that exploits this result to the greatest extent should be devel- ‘oped, so that additional computational benefits can be gained. Given the deficiency that stems from the lack ofa formal proof ate for the validity of this property, we did not develop a sophis- ticated design for the cost-matrix search, which explains the rather simplistic design that we present above, 6, SUMMARY AND FUTURE WORK In this article, we developed 1 mixed integer nonlinear ‘model for flight and maintenance planning of a group of aircraft that comprise a unit. The objective is to provide a balanced allocation of the flight load and the maintenance capacity to each individual aircraf, so that the long-term availablity of the unit is kept at a high and steady level ‘The formulation that we propose is based on a suitable mod- ification of an existing graphical heuristic too! for addressing this problem. Utilizing the problem's special structure, we also developed an exact search algorithm for its solution. ‘Our computational results demonstrate the superiority of the proposed algorithm over acommercial optimization package. Thisarticle points to several promising directions for future research. The question of whether the computational perfor- ‘mance of the proposed algorithm can be further improved through appropriate enhancements remains open. Therefore, investigating whether the special convexity result discussed in the previous section is valid stems as a very interesting task, Another interesting direction for future research is 10 apply the proposed medel on a time horizon of several time periods, to compare its long-term performance and behavior. ‘This will also render this model directly comparable to other, models from the existing literature, which are targeted toward longer planning horizons. Future articles should also be directed toward the devel- ‘opment ofa stochastic model that will incorporate the uncer- tainty that some of the problem's parameters might exhibit ‘Weedeliberately decided to deal withthe deterministic version Of this problem in this article, because itis novel, complex, interesting, and realistic enough to constitute a contribution in itself, which is useful for researchers and practitioners Moreover, the deterministic model that we address in this, article is a basic building block toward developing a more complicated stochastic model that will take into consider- ation random events, such as unforeseen failures. A recent example in that direction is the article by Mattila etal. [9] ‘who use diserete event simulation to model the maintenance of military aireratt in the Finnish Air Force and study its impact on aircrat availability bth under normal and conflict ‘operational conditions. Two factors of the problem under consideration exhibit significant uncertainty in practice: the fact that the actual duration of the maintenance service may turn out to be longer than its nominal value (expressed by parameter G), and the fact that a mission aircraft may fail in fulfilling its entire flight load (expressed by decision variable x,), The grounding ‘of an aiveraft for a longer time period than the one deter- ‘mined by parameter G may be dictated due to the detection Naval Research Logistics DOT 10.1002/naw28 Naval Research Logistics, Vol. 59 (2012) of serious findings during the regularly scheduled mainte- nance inspection or due to the lack of specific spare parts or Stalf expertise, which are needed to complete the service of the aircraft according to the prescribed safety standards. On the other hand, an aircraft may not beable ofl the entire time that has been assigned (o it, due to an unexpected event such san unforeseen failure. OF course, in case of such undesi able events, the user always has the option of reapplying the ‘proposed model after updating the state of the system, but a ‘Stochastic model would clearly incorporate such uncertainties ‘more accurately, leading to a better long-term performance for the entire system. APPENDIX. PROOF OF PROPOSITION 2: Let Ay, Ap. and (l= J...) be ‘te non-negative dal mokipios of constraints (24)-(26, eepctsel In ‘xkltion tothe original constants ofthe problem, the KK conditions a 2p — ks 4) — Ha 4a be > Iki G0) IKI Bb, aL-24p — Hg) Ay 2 Had Mu (= Ss) ° om he (Es -05+ =) =0 3) alte K=O =D oH dy20 dp20 m20, & x as) We cal the quantity (4p As 14) the “perpendicular distance of arcralt fom the diagonal” Despite he wo "itance” inthis inion, not that this quunty negative when arate Blow the diagonal tthe end ofthe eurent period. As LL. = UL always, there are 12 posible distin rangement ofthe quantities LL, UL, D. and X. When there does et exist fan arangement in which LL precedes X. the problem i clea infeasible [ecause the Might requirements (constant (2) cannot he satisfied. even ‘when every aircraft is asignod its manitum possible ight time, tn each ofthe romain eit cases, is clear hat the solution obained feo the application ofthe Procedure Sweep wes the sum ofthe assigned ical Tight dimes becomes equal othe second dusty in the reangemet,sats- fas (24) (26) and is there feasible. We show next tha ths solution also satisfies coulitons (30)-(3), and is thevefore optimal 00. (CASE 1: The arangomeat is (LL, UE, D.X} o (LL, UL.) In this ase, the sum ofthe assigned scat Hight tes inthe obained solution sequal othe second quantity inthe arangement, UL. We partion ‘he indices of the decison variables ofthis sclton ino four sts 4 st Sy contains the indices ofthe variables suc that xq = 0 Xs ‘set Sy coms the indices ofthe vaiales x sch hat xg = 0 X ‘set $5 comin the indices of he vaiales a such hat Oe = re 4. Se 4 contains he indices ofthe variables 2 such hat = re © Xs Westar 24onp — bs — nu) forsome k € 84 Noe that he valu of isthe same fr any & Si a eS contans the indies of the [Naval Research Logistics DOT 10.1002/nav ‘varales that ion the sweeping line atthe cute solution, anda result heir perpendicular distance from the diagonal isthe same. Additionally, each ofthese distances non-nogatve, because the ae hat UL appears before D inthe avagement imps thatthe sweeping Tin His above the Aiagoal at the caren solution; therefore 23 i non-negative, too, We aso sey = max(2(qny —k8 44) ~ 33.0) fore Siu, = Olork © S3USy date = 20iy bs — 4) hy Fork © Sy. Tho mulipio uy fore Sy re always non-negative, a et 3) conan the indices of the vail tha ‘ser inly sept and ater disengaged by the sweeping line, cause they reached thet upper bound thesfors the perpendicular distance ofeach of ‘hess points from he diagonal cannot he salle th he perpendicular i= tance from the diagonal of any point that ex on the sweeping line 3 he cureat solution. Fork 5; and € 5. constrains (30) and (31) ae clearly Satisfied. Constraints (30) ace clearly satisfied as an equality for Sx ‘hereon, constraints (31) re sais, oo, For 8, comstaits 31) ae ‘leuly sted and constrains (30) ae sisi if 2 > 24g — Rs — 0, ‘which i trac a st Ss contain the indices of the arabes that have ot thon swept by te line yet; therefore, their perpendicular distance for the iagonal cannot he ager than the perpendicular distance rom te dgonal ‘oF any point hat lies onthe sweeping Tine at he curent solution, Finally constants (5238) ae clay said, too. Hence, the cueeat solution Cogether with, 22.andug (k= I,K} asthe dual mulpis sashes the KKT conditions and is therefore optima Tse Sy iscampt.thon the above analysis remains the same, but ip neds tw be set equal to max(mases, 2104p —A8 —).0) [bath 83 and Si ate cmp then every deinen arable hs akon is maNimum possible ale ‘This implies dat UL = X, and this case redices 9 Case 4 through a8 appropiate earangeneat (CASE 2: The arangomeat is (LL, D, UX] 0 [LL D.X, UL) In this cae, de sn ofthe assigned sera fight ses in the sition ‘blaine qa tothesecond quant inthe arrangement, Dwhichimplics thatthe sweeping line coincide with the diagonal. We partition the indies ‘ofthe dino arial of his solution nth same four sets sin Case We set 4 = 2 = O.nk = maxi2(rip As ~14).0) fork € Six = Ofer S20 Se and ag = 204p ~ As — ra) fore Sy The multipin a foe [ee Ssareaiwaysnon-negalve ast Ss contains th indices othe arabes ‘hat wer inal swept and later disengaged bythe swecping line, because they reached ther upper bound thersfrs asthe sweeping ine coincides ‘withthe diagonal a the current solution, the perpendicular distance ofeach, ofthese ins mm he digonal cant he nepative. Fork © Sj and € 5s, ‘consrams (30) and 31) areal aise, Fork ¢ Sp, contains (31) are leu sted and constraints (30) te satsied if —2(j4p— As — 4) 20, ‘which is true as et Sp contain the indices of the arabes thar have ot ton swept hy the line ys therfore, as the swscing Fine coincides with the diagonal a the curea solution, cach ofthese points has now postive petpendculr distance fom the diagonal. Set, contains the indies of the ‘acabls tha eon the seeping line atte curentsalition. As the sweep ing ine coincides with the dagen, he perpendicular distance ofeach of ‘hese variables from the diagonal is qual oO. Asa result coms (30) sind GI) are also staid foe K © Sy. Finally, consents (32)435) ate ‘leuly stised, too. Hence, the cumen solution together with 4,33, and 14 =... [Aas the dal muller satis the KKT condions and is therefor opin. (CASE 3: The arangsment is (D: LL, UX] 01D. UL X,UL) In this case, the sm ofthe assigned aircraft gh times in the sation obtine i xual the second anti inthe serangement, LL. We pai jon the nos othe dessin vrs his sTaion no he sa Four sets a5 in Cases I and 2. We set ig = Oand 24 = ~20p ~ ks — m0). for & © Sg, Note thatthe valve of isthe same For any E € Su as se ‘St comin the ince ofthe variables that lie onthe sweeping line athe ‘urent soetion, and sa es their perpendicular distance from te dg ‘nals the same Addons ach of hese distances is nonspositve 35 the fact hat appears ist i the arrangement implies tat the sweepingKozaniis, Gavranis, and Kostaelou: Flight and Maintenance Planning of Mission Aircraft 29 line does nt leave the diagonal athe cuseat lution: therefore ison-negatve, oo, We also Seay = MC — AB — 4) + 1,0) foe Ee Shug = Ofork © 810 Sq and uy = 2p —ks —m4)-+2) fork © S ‘The muliplics ug fore S) are always now-acgative, asst Sy contains indices ofthe triable that were nally wey and itr disengage by ‘seeping line, because they reacad thee upper bound: therefore, the ‘perpendicular distance ofeach of hese pins om the diag cannot bo ‘srllr than the perpendicular distance frm the diagonal of any pin hat Tas om the sweeping line athe curen solution. For ke Sy aml K€ Ss, onsen (30) an (31) ae clay satis. Constraints (30) ae scaly ‘tnd an equality fork € Se therefore, constrains (81) ar aie, too. For ke 8 constraints (31) ate larly sts and constrains (30) ae satisiod if “21 2 pap ~ ks — 1). which i te as et S2 conan the indies ofthe wriabls that vena vn swept th ine yt therefore the ‘erpenicule distance ofeach of these pint rom te diag cannot Be Taygerthan the perpendicular distance from tb diagonal of any poi hat is fon the sweeping Tie atthe current solution. Filly, constrains (32)-35) fat clearly satisfied, to, Hence, the cute solution Loge wilh 20, andy (= I, [A asthe dal multiplies ass the KKT conditions fn is therefore opi set Ss isempy then the above analysis emis the sane, but A aes tobeset equal to minjess(~20p —£5—4)- which always aon-pegatve {st $y contuns the indies othe vriahes tht have not heen swept {he line ye: therefore asthe sweeping ine does ot is above the digoaal athe curet solton, ach ofthese pots has non-positive perpendicular ‘isuance Tom the diagonal I ah Soa tt emp. thn very decision ‘ariale has taken is maximum possible value. ID = LZ = this ase seduces to Case trough an appropeat rearrangement. WD = LL = X, he seeping ines helo the diagonal ath curent solaton Ih me, we partion the indies of the decision variables ofthe eurensliaion 1 st Sy contains the indices ofthe variables such thal Xs 1 st) comtans the indices ofthe variables sto that O < 39 = Xt Weset2y = —26y4p—fs—x)-whorok € Sp sthe index ofa vail thats ciety onthe sweeping ine, max Q(vip K-00) $21.0), Fork © Sy and ag = 2004p Ks) +, fork Sp. Ashe sweeping line lies below the diagonal is sel pine. Additonal. the muller 1 for F © Sp ae alays non-negative, Bocase the perpendicular distance fiom the diagonal of ay pot that es onthe swecping Hine tthe eueat soluion cannot be lager than th perpendicular distance from the diagonal ‘fang other point Constaits(30)-(38) ar clay satisied fork © S15. ence the erent solution together with 2.42, and ap (E= Ty) a8 ‘he dil multipass the KKT conditions ai is therefore opin (CASE 4: The aangomeat is (LEX, UL, D] o (HX. DUE), In this ease, te sum of dhe assign aircraft times in the solution ‘btn seal tothe second quantity ia the arangement,X. We partion rcs ofthe decision variables of his ston ina wo ss 1 st 5, comtins the inices ofthe variables sich that; = 0 Xi ‘sot 5 comin the infices ofthe vaisles x, sich hat <= x, = Xs We set Ay = 2 = 0, and ay = maxQGyy — ke — 0) 0) fork € 81 andy = 20 4p — &8 — 4) fork © SThe multiplies for ke Sp are always nan-aepative a6 et contin the indices ofthe variables ha have tea been swept bythe sweeping ine o tsi upp bound: here 8 {he swecping line docs aoe low the diagonal a he cureat sluon, ‘ei perpendicular distance from the digonal non-negative. Consrais (20-35 are cleat sted for 5) U Sy Hence the creat Solution, together with. 22.ande (& = I...) asthe dual multipliers, sass the KKT conditions andi therefore opti. a ACKNOWLEDGMENTS ‘The authors thank Prof. Liberopoulos of the Department of Mechanical Engineering of the University of Thessaly for several fruitful discussions that led to the problem for- ‘ulation of Section 3, Nikolas Kanonis, an undergraduate student of the same department, for helping with the numer- ical results, and an anonymous referee for his/her valuable comments. REFERENCES U1] MS. Bazarss, HLD. Sherali, and CM. Shetty, Noafinear programming: Theory and algorithms, Wiley-Interscieace, New York, 2006, [2] P. Brucker, An O(n) algorithm for quadratic knapsack prob- lems, Oper Res Lett 3 (1984), 163-166. [3] PH. Calamai and JJ. Moré, Quasi-newton updates with hounds, SIAM J Numer Anal 41987), 1434-144 [4] Ro Helgason, J. Kennington, and H. Lall, A. polynomially bounded algorithm fora singly constrained quadratic progrun, “Math Program 18 (1980), 338-343. [5] G. Kozanidis, A multiobjective model for maximizing Neet availability under the presence of flight and maintenance requirements, J Adw Transport 43 (2008), 155-182. [6] G.Kozanidis. A. Gavranis, and G. Liberapoulos, “Heuristics for maximizing feet availability subject to flight & mainte- nance requirements.” in: Proceedings ofthe 10th International Conterence on Applications of Advanced Technologies in ‘Transportation, Athens, Greece, 2008, 15 p. (71 G. Kozanidis, G. Liberopoulos, and C. Pitsitkas, Flight and maintenance planning of military aireraft for maximum feet availability, Miltary Oper Res 15 (2010), 53-73, {81 LINGO 1110, User's guide, LINDO Systoms, ine., Chicago, TL (2008), Available at: hiphvww lindo.comy. Accessed on 17 February, 2012. {9] V.Matila, K.Vitanen, and. Raivio, Improving maintenance decision making in the Finnish Air Force though simulation, Interfaces 38 2008), 187-201 [10] PM, Pardalos und N. Kovoor, An algorithm for singly con- strained class of quadruie programs subject to upper and lower bounds, Math Program 46 (1990), 321-228 11] BLW. Pippin, Allocating Hight hours to army helicopters, MSe ‘Thesis, Naval Postgraduate School, Monterey, CA, 1998. [12] N. Safsei, D. Banjevie. and A.KS. Jardine, Woskforce- ‘constrained maintenance scheduling for military sircraft let ‘Accase study, Ann Oper Res 186 2011), 295-316. [13] A.Sgasik, Planning german army helicopter maintenance and mmission assignment, MSe Thesis, Naval Postgraduate Schoo Monterey, CA, 1994. [14] A. Steiner," heuristic method forareraft maintenance sehed- tuling under various constraints,” in: Proceedings ofthe 6th Swiss Transport Research Conference, Ascons, Switzerland, 2006, 28 p. 115] US. DoA, Field manual No. 304.500: Army aviation maintenance (Appendix D: Maintenance management tool), US. Deparment of the Amy, Washington, DC, 2000, Available at: hiip/uwye globalsccurity orgmilitaryibracy! policyfarmy/inv3+04-5001. Accessed on 17 Febeuary, 2012. Naval Research Logisties DOT 10.1002/na