You are on page 1of 10
IEEE Transactions on Power Systems, Vol. 11, No.1, February 1996 A GENETIC ALGORITHM SOLUTION TO THE UNIT COMMITMENT PROBLEM SA. Kazartis AG. Bakirtzis (Member) V. Petridis (Member) Department of Electrical and Computer Engineering . Aristotle University of Thessaloniki, Greece Abstract: This paper presents a Genetic Algorithm (GA) ‘solution to the Unit Commitment problem. GAS are general purpose optimization techniques based on principles inspired from the biological evolution using metaphors of mechanisms. such 35-natural selection, genetic recombination and survival Of the ftest. A simple GA algorithm implementation using the standard crossover and mutation operators could locate near ‘optimal solutions but in most cases faled to converge to the ‘optimal soluton. However, using the Varying Quafty Function technique and adding problem specie operators, satsfactory solutions to the Unit Commitment problem were obtained. Test results for systems of up to 100 units and comparisons with results obtained sing Lagrangian Relaxation and Oynamic Programming are also reported. Keywords: unit commitment, genetic algorithms. ‘1.INTRODUCTION ‘The Unit Commitment (UC) problem in a power system Involves determining the start-up and shut down schedules of thermal unis to be used to meet foreoasted demand over @ future short term (24-168 hour) period. The objective is to rminemize total production costs while observing a large set of operating constraints. The UC problem is a. complex ‘mathematical optimization problem with both integer and Continuous varlables, The exact solution to the problem can be ‘obtained by complete enumeration, which cannot be applied to feallstle power systems due fo lis excessive computation time requirements (1. Research’ efforts, have, therefore, concentrated on ‘efficient, suboptimal UC algorithms which can be applied to realistic power systems and have reasonable storage. and ‘computation time requrements. The basic UC methods reported in the Iterature can be classified in five categories: Proety List Ra) Dynamic Programming [4-7] Lagrangian Relaxation [8-14] Branct-and-Bound 15-18) ‘© Benders Decompesiion [17-18 95 WH 152-9 PARE A paper reconnended and approved ‘by the THEE Power Syston Engineering Committee of the TERE Power Raginoering Society for presentation et ‘the 1995 TEEE/PES Winter Mooting, January 29, to February 2, 1995, Now York, WY, Manuscript subai ted aly 27, 19945 nade available for printing ‘Sanuary’3, 1995. 0885.s950)96/805. ‘Since improved UC schedules may save the Electric Utities milons of dotars per year in production costs, the search for closer to optimal commitment schedules continues. Recent efforts incide application of simulated annealing (1), ‘expert systems [20] and Hopfed Neural Networks [2-22] for the soluton of the UC problem. “This paper gives’ a Genetic Algorithm solution to the UC problem. Genetic Algorithms (GAs) were first proposed by Holand [23] and more recently reviewed and enhanced by Goldberg [24], Davis [25] and many others. GAs have been applled to many diverse areas such as function optimization, ‘system —identifeation and contro, image processing, combinatorial probloms [26], Artificial Noural Network topology eterminaton and rule based systems. In power systems GAs have been recenty applied for the solution of the economic optimum Ean 20 = 0 m0 wo ao ato Generations Figure 4. Simple GA average progrose of the best cchromosome's quality over 20 runs. ‘The additional set of operators consists of the operators escribed below: 1) Swap-window operator ‘This operator is applied to all the population genotypes with a probabity of 0:3, selects two arbitrary unis, U2 a "time window" of width w (ours) between 1..H and a random window positon between 1.(H-W) as shown in Fig. 5. Then the bts of the two units ut, u2 included in the window are exchanged. This operator acts lke a sophisticated mutation operator ; ss 8 Average best qaity zor aa Optimum 8 Total Unit Cost 8 8 | wo 0 ms Generations Figure 6. GA with problem specific operators = average progress of the best chromosome's quallly over 20 runs. ‘Addtionaly another set of operators was implemented to apply hiclmbing-tke techniques to the best chromosome of ‘every generation. This set incudes : ‘a) Swap-mutation operator. ‘This operator |S applied only to the best genotype of ‘every generation. For every hour of the scheduling period, It performs with equal probabilty one of two tasks ‘Lit selects two arbitrary units and exchanges thelr bis for the specific hour (exchanges their on-off states) oF 2 it selects a singe arbitrary unt and fips is corresponding bit {or the specific hour from '0' to“? and vice versa. If the performed task for a specific hour produces a better solution the solution is kept otherwise ti restored in iia! slate. This operator resembles the stochastic hil cimbing ‘method, ) Swap-window hill-cimb operator. hours This operator is also appied only'to the best genotype of i isso H every generation but with a probably of 0.3. it selects two arbitrary units ut, u2 and a "ime window" of width w (hours) 1 between 1.H. The time window is first considered aligned with ul+2 O00T im ‘the fst hour ofthe scheduing period as shown in Fg 7 3 — wed 1100 0S hours Swindow Units 1234567 iH || = i ute 2/[OTTT Figure 6. The Swap-Window operator. 3 oo w4 | TT indow-mutation operator Sa This operator is aso applied to all the population ‘genotypes with a probability of 0.3. It selects one ‘unit at ———— a random , a "time window" of wisth w between 1.H and random window position between 1.(H-w). Then it mutates all the bits included in the window turing all of them to 1s oF al of them to 0s. This operator can be considered as @ mul Polnt- mutation operator of continuous Bits. ‘The new enhanced GA scheme was tested on the benct-test problem and exhibited improved performance over the former simple GA scheme in finding solitons near the lobal optimum. The resuts are shown in Fig. 6 Comparing Figures § and 6 it s evident that the new GA scheme is much more efficient at the inal of the execution and finds good solutions “earlier” than the simple GA Figure 7. The Swap Window hil-cimb operator. ‘Then the bits of the two units ut, u2 included in the window are exchanged. The new solution is evaluated and I it IS otter itis kept otherwise It le restored in its iia stat, ‘Then the window is shifted one hour up and tho above procedure Is repeated unt the window reaches the last hour Of the schedullng period. This operator resembles the former ‘Swap-window operator but it» performs. multiple window swapping tests and it checks if a better solution Is produced. In other words the operator performs. local hilkclmbing ‘optimization concerning the commitment of two out of N une using a swapping mechanism, 5. SIMULATION RESULTS = {he sinter 1, 20, 40, 60,80 3 ‘and 100 unit systems. A base set of 10 units was initaly Ss] averages aty Rena ag n'a har Grand ns. Te et eel Stash lard sch sown fa tare nk pnt oun ore gol hase aici he dna ne mg by 21h len 2p citzesssssssssamaness | Silos sad wep he pecbene wa rae tis nal cans eave ot nied Be Wo he *) a mm | mend ceertone Table 1 Problem data for 1-u base U poten, fsre 5 WAR arcing ae wT PTs Ta To [RT OTS SPS tet Chomowar's any ovr 30 nn ‘The new GA scheme with the hil-cimbing operators was [Prin {haw = ‘0 | 20 | 20 = toad Sa aa ine ones be eee he nea Ff on Trae 0 ‘optimal solutions very close to the global optimum as shown in 2 (Siw) | 16.10 | 17.26 [16.60 | 16.50 |- 10.70 Fag. 8. Je (Siw2-n) [0.00048 [0.00031] 0.002 | 0.00211 [0.00398 ‘Another step taken was the implementation of a smooth fTin. up __(h 7 i 2 ] . au Galil eat toe ieee talee name feat 8 [6-6 | $s Sere elleniinio te Seach spre farsa | Wor | me soo | sso |oo0 Ba Relteson cooete he” cpaconen the faarast)| Som [oom [roo {a | Cohan poy rae ee) wees hoearg Estar et ost at a a ihbcqaccentox Se ommanon niente ait ete fst o TS ORT OS ORS TORT wen a [Pmax (MW) 80. 85. 55. 55. 55. ene aa [Pmin (MW) 20. 25 10 10 10 max: final value multiplier of constraint L* Sih) 370" | ano _s60_{—e6s_{ 670 the gansaton ner ne) (Bw | 26 277 a600 | orzr [277g Sim the Maxamum number of gonerations the GA is alowed [2__(S!MW=n] [0.00712 [0.0007] 0.00473 [6.00227|0.00173 wr to run. [min uo (hy 3 3 4 4 4 wae ae area te tr sowie to peaty & fence Tet 8 tt tt neat oS ct d bet ey RE rsarcerg| ao | eo [ao | ao] Beare as ae” Ratan ne Pavaitceme wo | geo {- es] a | iro ably Pods evant camy tec me paraatwe ts —[-2-1-o | 9. ony ‘gaiy “inclen tau hs. vayeg. seneh paigene laf off wrt ‘ote amp ebay ere ed ee eg sc can ———— Exes sro and set conepene be onan [3 spn] — Tr aaa ‘With this technique of the varying quality function the GA. 2 0 = oe tryna wee te wale pimew om f——ta $e 9 = —— = $1 = a = {09} [io 3 20 |, ca 760) 3 300 —| aol \ — oo 3 yy | _titteessseasennnnannnne ean 7 Langan laattn. agin (8) was ao ome | pets pon nar Sa etn er each © ao om woo | ilove ner nad dacs or bo Ga Geran Bs Nees gs be ck eee Symi Poa Auten (2) i cmp Figure 9. GA with varying penallios added: average progress of the best chromosome's quality over 20 runs. state enumeration was also developed and used to solve the {O-unit problem. As shown In Table ll, the solutions of the DP, the LR and the GA, for the 10-unit problem, are identical In othr test runs not reported here, the GA provided solitons leven better than the DP wih compte state enumeration, due to the dficalty of the DP to model time-dependent constraints sscussed in Section 3.2. For the larger problem sols the GA, solifons were compared only withthe solitons produced by the LR algorithm, as the time and capacity requirements of the DP algorithm with complete state enumeration are prohibitive for probioms of this seal. For every problem set the GA used the advanced operators and techniques described in the previous section. In ider to avod misteading results due to the stochastic nature of the GA, 20 runs were made for each probiem set, with feach run Starting with cfferent random populations. For a Specific problem set, everyone of the 20 rine was terminated at the same generation limit, the lim increasing withthe number of units. A run was considered successful ft ‘converged on a solution equal to or beter than that of the LR. algorithm, ‘Table 111. Simulation results for up to 100-unit systems. [Unis] DP Plage Ret Ch [soit] solution [success] best [worst [a erating Cost $f Se Cost $]“%e WO [sasaze] 65805] 60 —] 965825] 570032) 0.76 Zap — | rra0680| —75—|aezaal 7139069) 0.51 -40-|- [2258503] 90 |” 2251am| 250706 -0:36 '60_| = | 3304066] 100 [3576825] 5384252) 0.22. 0-[ = 4526022] 19014504039] 450791 0.7 “oo | ~-—[sa57277] 100 —[sa27aa7] sea7eia| 0.19 pr [ris [reap Gs [500 2 20] 1000, 738 '40_[ 2000. 2697. '60_|—3000- 5840, ‘60_| —4000- "0036" 700-| 5000. 5733 Average five (ee) 8292888 3 No of Units Figure 10, Scaling of the GA execution time with the number of units, » ‘The populaton size was 50 genotypes in all runs. In ‘general, when the population size increases, the number of ‘generations required by the GA to converge to the optimum ‘olution decreases. On the other hand, the CPU time required for the evaluation of a generation Increases almost lncarly ‘with the population size. The popuiation of 60 genotypes was Chosen, after several test runs concerning populations of 50, 75 and 100 genotypes, because it was sighty more efficient (le. t was faster in reacting the same soliton wth equal robabily) ‘The test results are shown in Table I. For the GA, both the best and worst solutons produced are reported together with their diference as a percentage of the best solution. As shown in Table I, for large systems (more than 60 units), the GA constanty outperforms the LR unit commitment. The GA average time reported concems CPU time on an HP Apolo 720 workstation ‘The scaing of the GA execution time withthe number of units to be commited, is shown in Figure 10. Analysis of the results prasented in Table Ill and Figure 10 shows thatthe GA ‘execution time increases in a quadratic way with the number ‘of units to be commited. 6. CONCLUSIONS ‘A Genetic Algoritim soliton to the Unit Commitment problem has been presented. It was necessary to enhance a Standard GA implementation with the addition of problem ‘pectic operators and the Varying Quality Function technique in order fo oblain salisfactory unit commitment solitons. A, basic advantage of the GA soluton isthe fexbity it provides Jn modeling Doth time-dependent and coupling constrains. ‘Another advantage Is that GAS can be very easly converted to work on paraliel computers. A disadvantage of the GAS is that, since they are slochastc optimization algorithms, the ‘optimality of the solution they provide cannot be guaranteed. However our results indcate thatthe diference between the Worst and the best GA-provided soliton is very small (0.74% at most), Another disadvantage of GA-UC algorithms Is their high execution ume. However with the progress. inthe hardware of paral computing both dsadvantages of the GA UG algorithms wil soon be eliminated. Ri Ices [1] Ad. Wood and BF, Wotenborg, Powar Ganeraton peraton and Control, 1284, Jon Wiley, New York. [2] Happ Hi, R.G. Johnson, Wal. Wight, “Lage scale hygro-hermal_unit_commiiment-method and results", IEEE Trans. on PAS, VO. PAS, pp. S738, 17 (3) Baldwin, Cul, KM. Dae, RE. Ditch, “A study of economic. ‘shutiown of generating units in daly poten", AIEE Tron PAS, VoL 76, 1260, pp 1272-288, (4) GA Pang’ and H.C. Chen, “Optimal Shor-Term ‘Thermal ‘Unt Commiment, IEEE. Transactions on Power Apparatus and Syston, VO. PAS-95, NO. 4, PP. “36-846, dulyAugst, 1576. 15) CKPang, GB, Sheble, and F, Abuyeh, “Evauaton of Dynamic Programming Based Methods. and. Mule ‘toa. Represoniaton fr Thermal Unt. Commitments" ‘EEE, Transactone on Power Apparatus and Vol PAS-00, No’, March, 36% pp. 122.128. @ (10) tm (2) rr) 4) (15) (6) ft) (6) (9) (20) ry (ea) WL Sryder, H.0. Powel, Jr, and J.C. Rayburn, “Dynamic Programming Approach to Unt Commament”, IEEE Transactions on Power Systems, Vol. PWRS-2, No. 2, May 1987, pp. 339-350, W.J. Hobbs, G. Hermon, S. Wamer, G8. Sheble, "An Enhanced Dynamic Programming Approach for’ Unit Commitment’, IEEE Transactions on Power Systems, Vol 3, No.3, August 1988, pp. 1201, 1205, Lauer, G.S,, DP. Bertsekas, NR. Sandel, J, TA. Posbergh, ""Soluton of large-scale optimal unit commitment problems", IEEE Tran. on PAS, Vol. PAS- 101, pp. 79-88, January 1982. ‘A.'Merin, and P. Sandrin, "A New Method For Unit Commitment At Electrcte De France", IEEE Transactions on Power Apparatus and Systems, Vol PAS=02, No. 5. Cohen, Al, SH. Wan, “A method for solving the fusl ‘constrained unit commitment problem’, EEE Trans. on Power. Vol, PWRS-2, pp. 608-614, Aug. 1967. Aoki, T. Satch, M. lio, "Unit commitment in a large- scale power system inchiding fuel constrained thermal and. pumped-storage hydro", IEEE Trans. on Power ‘Systems, Vol. PWRS-2, pp. 1077-1083, Nov. 1987. F. Zang, FD. Galana, "Towards a more rigorous and practcal unit commitment by Lagrangian relaxation’, IEEE Trans. on Power Systems, Vol. PWRS-3, pp. 763- 770, May 1988, FIN. Lee, "A. Fus-Constraited Unit “Commiment Method", IEEE Transactions an Power Systems, Vol 4, No. 3, August 1989, pp. 631-698. SVemani, C. Adrian, K.imhof, S. Mukherjee, “Implementation of @ Lagrangian Relaxation Basod Unit Commitment Problem’, EEE Transactions on Power Systems, Vol. 4, No. 4; October 1889, pp. 1373-1380. T.S. Dilon, "Integer Programming ‘Approach to the Problem of Optimal Unit Commitment with Probabiistic Reserve Determination’, (EEE Trans. an PAS. Vol PAS-A7, pp. 2154-7164, Nov./Dec. 1978. Al. Cohen, M. Yoshimura, "A Branch-and-Bound ‘Agoritim for Unit Commitment", (EEE Trans. on PAS, Vol. PAS-102, pp. 444-451, Feb. 1983. G. Cote, M.A. Laughton, “Decomposition Techniques in Power System Panning: The Benders Paritioning Method". Electrical power and energy systems, Vol. 1= Not= Apel 1979 LF. Baptstela, LC. Geromel, "A Decompostion ‘Approach to Problem of Unit Commitment Schedule for Hydrothermal Systems", IEEE Proc., Vol. 127, part D, No 6, p. 250, Nov. 1980) F Zhuang,” FD. Galina, "Unit Commitment by Simulated “Anneaing’, IEEE Transactions on Power ‘Systems, Vol. 5, No 1, February “980, op. 31-318. ©. Wang, SM. Shahigenpour, "A decompostion ‘approach to nonlinear multarea generation scheduling With te-ine constrants using expert systems", (EEE Transactions on Power Systems, Vol. 7, No. 4, November 1092, pp. 1409-1418, Z. Ouyang and SM, Shahidehpour, “A MutsStage inteligent™ System for Unit Commitment’, EEE ‘Transactions on Power Systems, Vol. 7, No.’ 2, May 1992, pp. 639-646 H. Sasaki, M, Watanabe, R. Yokoyama, “A. Solution Method. of Unit Commitment by Artificial Neural Networks", JEEE Transactions on Power Systems, VoL, No 3, August 992, pp. 974-981, JHlHoland, "Adaptation in Natural and tical ‘Systems", Ann Arbor, Michigan: The University of Michigan Press, 1975. DE. Goldberg, Genetic Algoritms in Search, Optimization and Machine Learning, Reading, Mass. 25]. Davis (ed) Hanabook of Genetic Algoitems, Van Nostrand, N.York, 1991 [26] _V. Petrdis and S. Kazaris "Varying Qualty Function in Genetic Algorithms and the Cutting Problem" to be published In the WCCI-94 (Evolutionary . Computation) Proceedings. A.Bakirtzis, V.Petrds, S. Kazars, “A Genetic Algorthm ‘Solon 9 the Economic Dispatch Problem" to be published in IEE Proceedings, Part C. D. Wallers, C. Sheble, B. Gerald, "Genetic Algorthm ‘Souton of Economic Dispatch with Valve Point Loading", IEEE Transactions on Power Systems, Vol. 8, No. 3, Aingust 1098, pp. 1325-1332. GB. Sheble, K. Billig, "Refined Genetic Algorithm ~ Economic Dispatch Example", paper 94 WM 199-0 PWRS, presented atthe IEEE Power Engineering ‘Society 1994 Winter Meeting, New Yor, ea) (24) BLOGRAPHIES ‘Soyos Kazars was bom in Thossloiki, Greece, ne 1866 Ho recoved the Dib. Erg. degree fom the Deparment of Becta Enger al the Asie Unversty of Thessaink, Greece in 1990. Shon 1060 ho has boon workng 19 @ Computer Anat, Pregranener land Lecter for pute and private compares. inne 10h ested Fb PhD at tho. Bctioa| Ergronring Depatment of ti Avsioto UUniveraty of Tresialni, Gresee, here he © pasar werk a6 @ foctors eandkatn He rosoarcy invests ‘ae. Evluorary Computaton (Garete’ Agere, Evauiorary Progamming ec), Aiitea"Newsl Newors, Sofware Engoeerrg” and” Computer “Techrsogy ‘We Kacais i a member of the Society of Professional Engneess of Grooce Anastacia Bate was born In So, Greece, in Fobruay 1056. He rocowod the Dis. Eng. degre fom the Deparment of BBectical Engnaorng atthe Natooa Techea Unveray of thor, Groves, in 79 ard the MSE-E an Pr. dogs fom Googe Trott of Technsogy, Attn, 684 and (864 respectvaly. He has worked (904) ae coraitant to. Souttom Company See "386 he Joined the facuty of the Elston Engneerng Depart at the Fretoto Unvorsty of Thesaaonic, Greece, where hes an assocte profeecor. He ragareh ‘norets ein power systam oporaton ard ono rola anaes aed arate eneay sources. ‘br Bakrtas is a member of IEEE and tho Socely of Professional Engrs of reer, Vasils Pati rocetved the domain Eoctrcal Engnoorng ‘tom tha Natoral Tocris Unversity mn Alors, Grese, n 00. He ‘blared the M.Sc. and PhD. dopres in lector and systems form Keres Calage, Unversty of London, 70 ar 1974 respec ‘He hs boon const of Naval Racer Contr n Groce, Direct of the Department of Eaeonies and Computer Engen, fe Vien Chaitnan cf tho Facaly of Scie ar Compute Engnserig the Artote Unveraty of Thessloni He tcurry professor h the Deparment of lactones and Computer Enger {fe Artote Unversty of Tessar, Groce, ‘Ato of thre tooks on control and mesaurment stems end cover ify research papers. He rosoafeh Ptorstsniclale_ conta ystems, inligent and autonomous systams, arf neral network, (inate agorinms, robles anda automaton. Diseussion A.J. Conejo, N. Jiménez Redondo, J.M. Arroyo (Gniversidad de Mataga, Mélaga, Spain): We wish to com :mend the authors for their contribution in providing a new approach to address the complex unit commitment prob- lem. We would like the authors to comment on the following 1, ‘The random generation of the individuals of the first generation and the crossover procedures proposed to senerate the individuals of a new generation systers- tically produce infeasibiities with respect to time (ramp limit, minimum up time, and minimum down, time violations) and demand (spinning reserve and demand constraint violations). Heuristie procedures ‘ean be used to eliminate or to reduce the quantity andl the level of the infeasibilities produced, while main- taining an important level of freedom in generating the new individuals. ‘This in turn may results in more eflicient genetic algorithm. Have the authors any experience in using the aforementioned heuristic procedures? 2. ‘The progressive penalty method used to produce fea sible individuals allows it to clearly distinguish be- tween feasible and infeasible individuals. However, the resolution to differentiate among feasible individ. uals is low if the penalty term is large. This may re~ sults in a slow progressing genetic algorithra, What fs the experience of the authors in that respect? Finally, we would like to congratulate the authors for their Interesting and well written paper. Manuscript received March 6, 1995. S.A. Kazarlis, A. G. Bakirtzis, V. Petridis, We thank the discussors for their interest in our paper and for their valuable comments and questions. In response to the issues raised by the discussors, we ‘comment the following : 1. The use of heuristic techniques. that transform an infeasible solution tothe nearest feasible one have been used {nthe pas 1] in order to alleviate the problem of dealing With complex constraints that result in a search space that is roslly covered by infeasibic solutions. To our knowicdge, two techniques have been usod so far : a) to transform an infeasible solution tothe nearest feasible one, replace the old genotype in the population with the transformed one and ‘evaluate the new genotype, and, bt transform an infeasible solution o the nearest feasible one, keep the eld genotype in the population, but evaluate the transformed genotype and assign the quality of the transformed genotype tthe original infeasible one a In case(a) no infeasible solution is allowed to participate in the population, Thus, all the genotypes that the GA will ver consider will be the feasible ones, In cases of search spaces where the feasible solutions cover a large portion of the space, this may help, by avoiding the bad solutions. But in cases where the search space is mostly covered by infeasible solutions, as in the UC problem, the search is limited only to some of the subspaces of feasible solutions, reducing the diversity of the genetic material in the popula- tion and thus resulting in the possibilty of missing the arca ‘of the global optimum, as the subspaces of feasible solutions ‘may be isolated from each other and "far" from each other ‘concerning the operators’ recombination abilities. In case(b)the diversity in the population is maintained as, the infeasible solutions are not really transformed, but the assignment of the quality of the nearest feasible solution to the infeasible one, can be misleading to the GA, as, ‘depending on the heuristic method of the transformation, the transformed feasible solution that is evaluated can be very different from the infeasible one and thus lead the GA away from the global optimum (that does not need transformations tobe feasible) and towards infeasible solutions that transform to feasible ones. ‘The authors have used versions of the above mentioned {echniques inthe initial GA tests but the results where not as good as those obtained by the use of the Varying Quality Function technique, so these techniques were not included in the final version of the GA that was used to produce the reported results. 2. During the last generations of the GA, when the Constraint penalty factors obtain ther final large values, a Jarge proportion of the population consists of feasible solutions. The qulites ofthe feasible solutions are, say, in the range 102.103 and the qualities of the infeasible solutions may be inthe range 108 10. The selection method used to select parens for reproduction from within the population, is the Rouleue Wheel parent selection Inechanism that selects a genoype with 2 probability Proportional to its relave’ fitness in the population Considering Figure 1, the infeasible solutions will occupy @ ‘ery small proportion of the wheel's perimeter, while the feasible ones wil occupy the larger proportion ofthe whee. Pease note thatthe portion of the wheel assigned 10 a ‘genotype is proportional tits inverse ness as we del with 2 minimization problem. So, because of this relativity inthe Selection mechanism, the resolution between feasible Solutions is adequate forthe GA t0 distinguish the best Solutions among the feasible ones “Moreover, in order to enhance the selection pressure, on the bast of the genoypes, we have used a fitness scaling ‘mechanism, also described in the paper. The mechanism is ted in the cae of excosive diversity in the popalation, in oder to asist the convergence af the GA, by selecting more frequently the best of the genotypes dus 1 it technically increased fitness. When applied, the fness scaling imechanism increases the fines (fat the inverse fess for a minimization problem) of the genotypes that have & Infeasible solutions Feasible solutions 1. The Roulette Wheel parent selection fitness within 1% close to the fitness of the best genotype of the population, by a factor of 10. When the observed diversity that triggered the mechanism ceases to exis, the mechanism is withdrawn and all the fitnesses return. to normal scaling References [1] Ryobei Nakano and Takeshi Yamada, "Conventional Genetic Algorithm for Job Shop Problems", proceedings of ICGA'S1, pp. 474-479, Manuscript rossived April 19, 1995.

You might also like