ANÁLISE DOS RESULTADOS DO PAPER DISTRIBUTED BLOCKING FLOW SHOP
Hugo Hissashi Miyata
Prof. Dr. Marcelo Seido Nagano
1. Small instance problems IG_NM(3) and IG_NM(4) outperform MILP
The small instance problems set in quality of solution. In general, all includes all instances with F = {2, 3, 4} metaheuristics tends to increase the RPD factories, n = {4, 6, 8, 10, 12} Jobs, m = {2, with the expansion of number of jobs, 3, 4, 5} machines, l = {1, 2, 3, 4} while success rates tends to reduce. This maintenance levels and setup rates with trend is more expressive with IG_RPN(2), {10%, 50%, 100%, 125%} of the processing IG_RPN(3) and IG_RPN(4), although they times. are the methods with the lowest average In the average relative RPT among the methods. Also, the percentagem deviation (RPD), one can computational time consumed by the notice that HFFO(4), HFFO(3) and HFFO(2) referred metaheuristics seems to be stable outperform all remaining methods. regarding the number of jobs. Conversely, Conversely, in the mean success rate (SR), one can notice increase of RPT with n MILP outpperforms all other heuristic varitaion, since HFFO and IG_NM are methods evaluated, reaching the best structured with local Search and solution for approximately 77,7% of the reconstruction mechanism more complex small instance problems set. Regarding than IG_RPN. Regarding the number of mean relative percentagem deviation machines, one can observe inverse effects computational times (RPT), IG_RPN between MILP and the metaheuristics. algorithms, IG_RPN(2), IG_RPN(3) and With the increase of m, the metaheuristics IG_RPN(4), obtained the lowest values of tends to obtain reductions of RPD, while ARPT. On the other hand, IG_NM(3), MILP tends to increase. In general, success IG_NM(4) and IG_NM(2) perform closer to rates of HFFO, IG_NM and IG_RPN MILP, with ARPD and ARPT lower than algorithms tends to be stable with the IG_RPN and HFFO algorithms, respectively. increase of m, while SR of MILP reduces. Regarding the number of factories However, MILP outperforms all evaluated (F) parameter, one can observe that methods regarding SR. Efficiency is affected HFFO(3) and HFFO(4) outperform all by the number of machines, since RPT of all evaluated methods regarding ARPD. methods increase with larger values of m. However, one can notice that regarding SR, Since larger number of maintenance levels one can notice that MILP outperform all (l) becomes the problem more complex, methods. IG_NM algorithms obtain one can observe that RPD of all methods reductions of RPD with the increasing of F tends to increase. Although all versions of parameter, outperforming MILP in instance HFFO outperform MILP in average RPD, problems with 3 and 4 factories. Regarding one can observe that MILP outpeforms all efficiency, one can observe reductions of methods regarding success rates in all RPT of all metaheuristics with the increase values of l. In the mean, IG_NM(3) and of F. IG_NM(4) outperform MILP in instances Since the problem becomes more with l = 3 and l = 4. Regarding efficiency, complex with the increase of parameter one can observe that RPT of all methods values, MILP tends to reach the CPU time tends to increase with expansion of limit from 10 jobs on. One can observe number of maintenance levels. Regarding more difficults of MILP to solve problems the setup duration rates, one can notice with F = 2 and 3. In problems with 12 jobs, that RPD of all methods tends to increase HFFO(2), HFFO(3), HFFO(4), IG_NM(2), with larger values of setup duration rates. In the average RPD, IG_NM(4), HFFO(2), observe that RPT of all methods tends to HFFO(3) and HFFO(4) outperform MILP reduce with the increase of F. HFFO(4) considering all setup rate values. IG_NM(3) presented the largest RPT followed by outperform MILP in instances with setup HFFO(3) and HFFO(2). rates equals to 10%, 100% and 125% of the Regarding the number of jobs (n), processing times. Concerning success rates, one can observe that while RPD of one can notice that MILP outperforms all HFFO(2), HFFO(3) and HFFO(4) obtain the remaining methods, followed by HFFO, increasing values of RPD, while RPD of IG_NM and IG_RPD algorithms. The IG_NM(2), IG_NM(3), IG_NM(4), addition of different setup rate values IG_RPN(2), IG_RPN(3) and IG_RPN(4) tend seems not to alter the computational time to reduce with the increasing of n. One can consumed by the heuristics. On the other notice that HFFO(4) outperforms all other hand, an increase in RPT can be seen for methods for n <= 20 regarding average RPD MILP with larger values of setup rates. and average SR. For the remaining values of n, one can see that IG_NM(3) 2. Taillard’s instance problems outperform all other heuristics evaluated. In the mean, IG_NM(3), HFFO(4) In general, all evaluated methods tends to and IG_RPN(4) outperform their respective obtain increasing values of RPT with larger versions. One can observe that IG_NM(3) values of n. In the mean, IG_RPN outperforms all evaluated methods, algorithms obtain the best values of RPT, improving the initial solution given by although they generated the largest values EHPF2_B in aproximadately 9,69% of RPD. of RPD. Among all versions evaluated, Also, the referred method obtained the IG_RPN(4) outperforms IG_RPN(3) and best mean success rate (SR) among the IG_RPN(2) in both ARPD and ASR, while evaluated methods, followed by IG_NM(4) IG_RPN(2) is the most eficient version and IG_NM(2). HFFO(4) and IG_RPN(3) evaluated. obtained the best values of mean SR Concerning the number of among their respective other versions machines, one can observe that both RPD evaluated. Regarding efficiency, one can and RPT of all methods tends to reduce notice that IG_RPN(2), IG_RPN(3) and with the increase of m. Regarding RPD, one IG_RPN(4) are the fatest heuristics. can see that IG_NM(3) outperforms other Additionally, one can observe that, in the methods in all values of m, HFFO(3) and mean, IG_NM(3) is faster than IG_NM(4), HFFO(4) outperform IG_NM(2) for HFFO(4), HFFO(3) and HFFO(2). IG_NM(2) instances with m = 2 and HFFO(2) consumes lower computational time than outperforms IG_RPN(4), IG_RPN(3) and IG_NM(3), but, it generates a larger RPD. IG_RPN(2). Taking into account the success Regarding the number of factories rates, one can see that IG_NM algorithms (F), one can observe reductions of RPD of outperform all versions of HFFO and IG_NM(2), IG_NM(3) and IG_NM(4) with IG_RPN in all values of m. Efficiency is also the increase of F while RPD of HFFO(2), affected by the number of machines, since HFFO(3), HFFO(4), IG_RPN(2), IG_RPN(3), all algorithms consumes larger values of IG_RPN(4) and EHPF2_B increase. HFFO computational time to return solutions. algorithms obtain the best performance for Since each iteration of HFFO requires to instances with F = 2, while IG_NM evaluate all individuals of the population, outperform all theuristics in the remaining one can observe that the referred method F values. In all values of F, one can notice is the most affected by the increasing of all that IG_NM(3) outperforms all other parameters. methods in the average SR, folllowed by Regarding the number of IG_NM(4). Regarding SR values, among maintenance levels, one can observe that other algorithms, HFFO(4) and IG_RPN(4) RPD and RPT tends to increase with larger outperform their other respective versions. values of l. Different from what occurs with Concerning computational time, one can F, n and m parameters, performance of HFFO and IG_RPN are more impacted than IG_NM with the increase of l. On the other hand, values of SR of all methods tend to be stable with ascending values of maintenance levels. One can observe that SR of IG_NM algorithms outperform HFFO and IG_RPN in all values of l. Computational times tends to increase proportionally with larger values of l, significantly impacting the efficiency of HFFO and IG_NM algorithms, while IG_RPN has RPT increased with slight increments. One can observe that RPD of all methods tends to increase with larger values of setup rates. In general, IG_NM(3) and IG_NM(4) present the lowest means of RPD for each setup rate value evaluted. HFFO(4) and IG_RPN(4) outperform their respective other versions in terms of RPD. Concerning success rates, one can observe that IG_NM(4) outperforms its respective versions IG_NM(3) and IG_NM(2) as well as the remaining methods. HFFO(4) and IG_RPN(3) obtained the best average SR among their respective other versions. One can notice that setup rates variation has a little impact concerning the computational time of all methods, where the values of RPT of all methods is concentrated in the mean.