* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download An Amended Genetic Algorithm for Job Scheduling on Parallel
Survey
Document related concepts
Transcript
A Genetic Algorithm with Sub-Indexed Partitioning Genes and its Application to Production Scheduling of Parallel Machines Chichang Jou Department of Information Management, Tamkang University Motivation and Objectives Goal of Production Scheduling To seek optimal combination of short manufacturing time, stable inventory, balanced human and machine utilization rate, and short average customer waiting time Motivation Scheduling requirement from an electronics factory Investigation of scheduling algorithms Objectives A solution based on genetic algorithms A new approach in solving production scheduling of parallel machines Chichang Jou 2 Background and Scope Scheduling Classification Products: single products / multiple products Orders: make-to-stock / build-to-order Production: job shop / flow shop Stages: single stage / multiple stage Machines: single machine / multiple machine Scope multiple products, build-to-order, job shop, multiple stage, multiple machine (parallel) A NP-hard problem Chichang Jou 3 Related Work Traditional Optimization Time consuming and easily trapped in local optimum Not applicable to real production Heuristic rules Chichang Jou SPT (shortest processing time) EDD (earliest due date) SLK (least slack time) Advantages:good speed and easy to understand Disadvantages: hard to find acceptable solutions in some cases 4 Related Work Approximation Algorithms Simulated Annealing, Hill Climbing, Tabu Search, Genetic Algorithms, etc. All of them are Probability based Having a fitness function to evaluate the effectiveness of a solution Gradually approaching the optimal solution step by step Having a method to jump off local optimum Advantages of Genetic Algorithms Chichang Jou Evolution mechanism ----- crossover, mutation, reproduction 5 Our Differences Close to real production Adjustable beginning date, due date, earliness cost, tardiness cost Introducing machine setup cost, machine idleness cost into fitness function New chromosome design for complete solution space Modification of the algorithm bounce, select Chichang Jou 6 Working Environment of the Electronics Plant Products: power supplies, motherboards, and barebones Supported with an MRP system Jobs need to flow through workstations of SMT, DIP, and PCG machines, with 4, 3, and 3 parallel assembly lines respectively Production plan has been continuously interrupted by cancelled orders or new higher priority jobs Manual scheduling had to be performed based on previous experiences Chichang Jou 7 Expected Workflow of Production Jobs Chichang Jou 8 Introduction to Genetic Algorithms chromosome gene flow chart search space Fitness value Search space Chichang Jou 9 Design of Chromosome Suppose there are n jobs to be assigned to m machines. A chromosome is modeled by a sequence of the following n+m-1 distinct genes: Example: Chromosome: 1 2 *2 3 *1 4 5 Jobs 1, 2 are assigned to machine 1 Jobs 3 are assigned to machine 2 Jobs 4, 5 are assigned machine 3 Chichang Jou 10 Fitness Functions Fitness Function For each schedule s F (Cs , d s ,U s , Rs ) [ i ( Esi ) i (Tsi )] U s Rs i Esi (Csi , d i ) max( 0, Lsi ) Tsi (Csi , di ) max( 0, Lsi ) Chichang Jou 11 Initial Solutions Example: Chichang Jou 12 Evolution Mechanisms Crossover Chichang Jou 13 Evolution Mechanisms Mutation: Chichang Jou 14 Evolution Mechanisms Reproduction: Chichang Jou Just copy the chromosome of the parent to its next generation Ensure no worse next generation Good results may be generated by crossover between different generations 15 System Implementation Main User Interface: A. Job table C. Performance indicators Chichang Jou B. Result schedule D. Related supporting functions 16 Experimental Parameters job priority earliness cost function low i ( E si ) 2 medium high Chichang Jou tardiness cost function i (Tsi ) 5 i ( Esi ) 2 Esi 2 i (Tsi ) 5Tsi 5 i (Esi ) 2 Esi2 2 Esi 2 i (Tsi) 5Tsi2 5Tsi 5 17 Experimental Results Cross comparison of our results using our fitness function with those with Cheng’s fitness function Using Our Fitness Function allocating genes distinguishable identical 1.000 1.004 2.02% 1.91% best fitness function value 119 120 worst fitness function value 127 127 ratio of average fitness function value (standard deviation)/ average fitness function value Chichang Jou 18 Experimental Results Using Cheng’s Fitness Function Chichang Jou allocating genes distinguishable identical ratio of average fitness function value 1.000 1.049 (standard deviation) / average fitness function value 5.39% 6.46% best fitness function value 120 121 worst fitness function value 137 151 19 Experimental Results Comparison of Fitness Functions whether consider machine setup and idleness costs total processing time total tardiness time total idleness time No 311.810 10.700 27.810 Yes 307.200 10.410 23.220 Chichang Jou 20 Experimental Results Two additional steps: selection selection of good chromosomes ranking probabilistic choice bounce detection of premature convergence bounce Chichang Jou 21 Experimental Results Modified Algorithm: initial select chromosomes with ordered probability mutation crossover reproduction rank fitness function values reach max. generation? bounce the best solution yes no no premature convergence? yes record the best solution Chichang Jou 22 Experimental Results Effects of the two additional steps: normal genetic algorithm add selection step add bounce step add selection, bounce steps average fitness function value 1.12 1.03 1.04 1.00 standard deviation of fitness function values 21.18 5.07 2.55 2.11 (standard deviation) / average fitness function value 15.33% 4.02% 2.00% 1.71% Chichang Jou 23 Conclusions Our Characteristics Closer to the optimal solution than other genetic algorithms Small deviation of solutions in each run More suitable for real production with realistic considerations Chichang Jou 24 Future Research Topics Defining new modules of genetic algorithms Improving the structure of the genetic algorithms Other application of genetic algorithms Chichang Jou 25