Particle Swarm Optimization Algorithm for Unrelated Parallel Machine Scheduling with Release Dates

Abstract

We consider the NP-hard problem of minimizing makespan for 𝑛 jobs on 𝑚 unrelated parallel machines with release dates in this research. A heuristic and a very effective particle swarm optimization (PSO) algorithm have been proposed to tackle the problem. Two lower bounds have been proposed to serve as a basis for comparison for large problem instances. Computational results show that the proposed PSO is very accurate and that it outperforms the existing metaheuristic.

1. Introduction

This research considers the problem of scheduling 𝑛 jobs on 𝑚 unrelated parallel machines in the presence of release dates. The performance measure, makespan, is defined as max(𝐶₁,…,𝐶_𝑛), where 𝐶_𝑗 is the completion time of job 𝑗. Minimizing makespan not only completes all jobs as quickly as possible but also is a surrogate for maximizing the utilization of machines. Following the three-field notation of Graham et al. [1], we refer to this problem as 𝑅|𝑟_𝑗|𝐶_max. This problem is NP hard [2].

Chen and Vestjens [3] used the largest processing time (LPT) to minimize makespan for identical parallel machines with release dates (𝑃|𝑟_𝑗|𝐶_max). The release date of a job is not known in advance, and its processing time becomes known at its arrival. Kellerer [4] proposed algorithms for the 𝑃|𝑟_𝑗|𝐶_max problem and 𝑃|𝑟_𝑗|𝐶_min problem. Koulamas and Kyparisis [5] considered uniform parallel machine scheduling problems (𝑄|𝑟_𝑗|𝐶_max). They proposed a heuristic and derived a tight worse-case ratio bound for this heuristic. Centeno and Armacost [6] showed that the LPT rule performed better than the least flexible job (LFJ) rule for the problem with machine eligibility restrictions (𝑃|𝑟_𝑗,𝑀_𝑗|𝐶_max). Lancia [7] applied a branch-and-bound (b & b) procedure to solve scheduling problems with release dates and tails on two unrelated parallel machines (𝑅₂|𝑟_𝑗,𝑞_𝑗|𝐶_max). Similarly, Gharbi and Haouari [8] also presented a b & b procedure to solve the 𝑃|𝑟_𝑗,𝑞_𝑗|𝐶_max problem. Carlier and Pinson [9] reported new results on the structures of Jackson’s pseudopreemptive scheduling applied to the 𝑃|𝑟_𝑗,𝑞_𝑗|𝐶_max problem. Li et al. [10] used a polynomial time approximation scheme for scheduling identical parallel batch machines (𝑃|𝑟_𝑗,𝐵|𝐶_max). Li and Wang [11] proposed an efficient algorithm for scheduling with inclusive processing set restrictions and job release times (𝑃|𝑟_𝑗,𝑖𝑛𝑐𝑙.𝑝𝑟𝑜𝑐.𝑠𝑒𝑡𝑠|𝐶_max).

To the best of our knowledge, no research has yet been published that develops an efficient algorithm to minimize makespan for unrelated parallel machines with release dates. The rest of this paper is organized as follows. Section 2 presents our proposed lower bounds. Section 3 presents the proposed PSO. In Section 4, the computational results are reported. Section 5 presents our conclusions and suggestions for future research.

2. Lower Bounds to 𝑅|𝑟_𝑗|𝐶_max

We propose two straightforward and easily implementable lower bounds for the studied problem. 𝐿𝐵₁ is the maximum value of each job’s release date plus the minimum processing time (across all machines). 𝐿𝐵₂ is set to the minimum release date (among all jobs) plus the sum of all jobs’ minimum processing times (across machines) divided by the number of machines. We set lower bound LB equal to the maximum value of 𝐿𝐵₁ and 𝐿𝐵₂: 𝐿𝐵₁=max_{1≤𝑗≤𝑛}𝑟_𝑗+min_{1≤𝑖≤𝑚}𝑝_𝑖𝑗𝐿𝐵₂=min_{1≤𝑗≤𝑛}𝑟_𝑗+∑^𝑛_𝑗=1min_{1≤𝑖≤𝑚}𝑝_𝑖𝑗𝑚LB=max𝐿𝐵₁,𝐿𝐵₂. (1)

To illustrate the proposed lower bounds, we consider example 1, which has 2 machines and 7 jobs. The matrix of processing times for example 1 is given in Table 1: 𝐿𝐵₁{}=max27+15,4+29,3+40,9+32,13+31,17+8,1+44=45𝐿𝐵₂()=1+15+29+40+32+31+8+442=1+1992{}=100.5LB=max45,100.5=100.5. (2)

Table 1 

The matrix of processing times for example 1.

3. The Proposed PSO Algorithm

PSO was first introduced by Kennedy and Eberhart [12] for solving continuous nonlinear function optimization problems. PSO is based on the metaphor of social interaction and communication in flocks of birds or schools of fish. In these groups, there is a leader (the one with the best performance) who guides the movement of the whole swarm. In a PSO, each individual is called a “particle,” and each particle flies around the search space with some velocity. In each iteration, a particle moves from its previous location to a new location at its newly updated velocity, which is calculated based on the particle’s own experience and the experience of the whole swarm.

A population of 𝑀 particles are assumed to evolve in an 𝑁-dimensional vector search 𝑅^𝑁 such that each particle 𝑘 is assigned the position vector 𝑋^𝑡_𝑘=(𝑥^𝑡_𝑘1,𝑥^𝑡_𝑘2,…,𝑥^𝑡_𝑘𝑁) and velocity vector 𝑉^𝑡_𝑘=(𝑣^𝑡_𝑘1,𝑣^𝑡_𝑘2,…,𝑣^𝑡_𝑘𝑁), where 𝑥^𝑡_𝑘𝑑 represents the location and 𝑣^𝑡_𝑘𝑑 represents the velocity of particle 𝑘 in the dth dimension of the search space at the tth iteration and 𝑘∈{1,2,…,𝑀}, 𝑑∈{1,2,…,𝑁}. Each particle knows its position and the corresponding objective function. The local best position for each particle 𝑘 is encoded in the variables 𝑃^𝑡_𝑘=(𝑃^𝑡_𝑘1,𝑃^𝑡_𝑘2,…,𝑃^𝑡_𝑘𝑁), and the global best position among all particles is encoded in the variable 𝑃^𝑡_𝑔=(𝑃^𝑡_𝑔1,𝑃^𝑡_𝑔2,…,𝑃^𝑡_𝑔𝑁). The standard PSO equations can be described as follows [12]: 𝑣^𝑡+1_𝑘𝑑=𝑤𝑣^𝑡_𝑘𝑑+𝑐₁𝑟₁𝑝^𝑡_𝑘𝑑−𝑥^𝑡_𝑘𝑑+𝑐₂𝑟₂𝑝^𝑡_𝑔𝑑−𝑥^𝑡_𝑘𝑑𝑥^𝑡+1_𝑘𝑑=𝑥^𝑡_𝑘𝑑+𝑣^𝑡+1_𝑘𝑑, (3) where 𝑤 is the weight that controls the impact of the previous velocities on the current velocity, 𝑐₁ is the cognition learning factor, 𝑐₂ is the social learning factor, and 𝑟₁ and 𝑟₂ are random numbers uniformly distributed in [0,1].

PSO has been successfully applied to a variety of continuous nonlinear optimization problems. In recent years, considerable effort has been expended on solving scheduling problems by PSO algorithms. Articles [13, 14] used PSO algorithms to solve scheduling problems similar to the problems in this paper. Reference [13] provided a PSO algorithm for scheduling identical parallel machines to minimize makespan (𝑃||𝐶_max). Reference [14] presented a PSO algorithm for scheduling nonidentical parallel batch processing machines to minimize makespan (𝑃|𝑏𝑎𝑡𝑐ℎ,𝑠_𝑗,𝑆_𝑚|𝐶_max). This research uses PSO for the 𝑅|𝑟_𝑗|𝐶_max problem. The following five headings describe the PSO algorithm used in this research: particle representation, initial population generation, particle velocity and sequence metric operators, local search, stopping criteria, and parameter settings.

3.1. Particle Representation

A coding scheme developed in [15] is used to represent a solution to the problem at hand. The coding scheme uses a list of job symbols and partitioning symbols. A sequence of job symbols, denoted by integers, represents a possible sequence of jobs. The partitioning symbol, an asterisk, designates the partition of jobs to machines. Generally, for an 𝑚-machine 𝑛-job problem, a solution contains 𝑚−1 partitioning symbols and 𝑛 job symbols, resulting in a total size of (𝑚+𝑛−1). For example, for a schedule with 7 jobs and 2 machines, the particle can be represented as shown in Figure 1.

Figure 1 

Representation of a particle’s solution.

The completed schedule is thus jobs 7, 6, 1, and 4 on machine 1; jobs 3, 2, and 5 on machine 2. This coding scheme specifies not only which jobs are assigned to which machine but also the order of the jobs on each machine. These pieces of information are important, since we are scheduling unrelated parallel machines with release dates.

3.2. Initial Population Generation

In order to give the PSO algorithm good initial solutions and to increase the chances of getting closer to regions that yield good objective functions, we propose a heuristic, named SRD_Reassign. The proposed heuristic SRD_Reassign is described as follows.

3.2.1. Heuristic SRD_Reassign

Step 1. Let 𝑈 be the set of unscheduled jobs; let 𝑡_𝑖 be the sum of the processing times of the jobs that have already been scheduled on machine 𝑖, 𝑖=1,…,𝑚. Initially, set 𝑈={1,…,𝑛} and 𝑡_𝑖=0, for 𝑖=1,…,𝑚.

Step 2. Arrange the jobs in the order of the shortest release date (SRD) first, and then assign the job to machine 𝑖^∗ that has the minimum processing time, that is, 𝑝_𝑖∗𝑗=min_{1≤𝑖≤𝑚}𝑝_𝑖𝑗. Repeat until all jobs have been scheduled to generate a complete schedule.

Step 3. Let 𝐴_𝑖 be the set of scheduled jobs on machine 𝑖, 𝑖=1,…,𝑚; let 𝐶_max=max_{𝑖=1,…,𝑚}{𝑡_𝑖} represent the maximum completion time and denote the set of candidate jobs for reassignment as 𝐵. Initially, set 𝐵={𝑛𝑢𝑙𝑙}.

Step 4. Identify machine 𝑖 for which 𝑡_𝑖=𝐶_max. For every job 𝑗, 𝑗∈𝐴_𝑖, search for machine ℎ(ℎ≠𝑖), such that if job 𝑗 was reassigned to machine ℎ and the jobs on machine ℎ were sorted in SRD, the new calculated 𝑡_ℎ would be smaller than 𝐶_max, that is, 𝑡_ℎ=max(𝑡_ℎ,𝑟_𝑗)+𝑝_{ℎ𝑗,𝑗∈𝐴ℎ}<𝐶_max. If job 𝑗 and machine ℎ can be found, update the candidate set by adding job 𝑗 on machine ℎ to the candidate set 𝐵 by setting 𝐵=𝐵∪{(ℎ,𝑗)}. If 𝐵={𝑛𝑢𝑙𝑙}, then go to Step 6.

Step 5. Select machine ℎ and job 𝑗 from 𝐵 for the reassignment that has maximum Gain, where Gain=max{𝐶_max−𝑡_ℎ}. Reassign job 𝑗 to machine ℎ by setting 𝐴_ℎ=𝐴_ℎ∪{𝑗}. Sort the jobs on machine ℎ in SRD and update 𝑡_ℎ=max(𝑡_ℎ,𝑟_𝑗)+𝑝_{ℎ𝑗,𝑗∈𝐴ℎ}. Remove job 𝑗 from machine 𝑖 by setting 𝐴_𝑖=𝐴_𝑖⧵{𝑗}. Sort the jobs on machine 𝑖 in SRD and update 𝑡_𝑖=max(𝑡_𝑖,𝑟_𝑗)+𝑝_{𝑖𝑗,𝑗∈𝐴𝑖}. Set 𝐶_max=max_{𝑖=1,…,𝑚}{𝑡_𝑖} and 𝐵={𝑛𝑢𝑙𝑙}. Go to Step 4.

Step 6. Terminate the procedure.

The first two particles are generated by first-come, first-serve (FCFS) rule and SRD_Reassign. The remaining particles are generated by applying local search to the solution found by SRD_Reassign. For FCFS rule, we consider all unscheduled jobs and schedule each one on the first available machine according to FCFS. Local search is done by randomly choosing two jobs 𝑗₁ and 𝑗₂ from the solution found by SRD_Reassign and then interchanging jobs 𝑗₁ and 𝑗₂ to generate a new solution. From the initial population pool, we identify the best current solution 𝐶^∗_max and update the global best location (𝑃^𝑡_𝑔).

3.3. Particle Velocity and Sequence Metric Operators

Kashan and Karimi [13] worked on the classical PSO equations to provide a discrete PSO algorithm that maintained all major characteristics of the original continuous PSO equations when solving parallel machine scheduling problems. In this research, we use the two equations proposed in [13] to update the particle velocity and the sequence metric operators as shown in (4) and (5): 𝑉_𝑘^𝑡+1=𝑉O^𝑡_𝑘⁺𝑅O₁^×𝑃^𝑡_𝑘O𝑋O^𝑡_𝑘⁺𝑅O₂^×𝑃^𝑡_𝑔O𝑋^𝑡_𝑘𝑋_𝑘^𝑡+1=𝑋O^𝑡_𝑘⁺𝑉_𝑘^𝑡+1. (4)(5) In (4), 𝑉^𝑡_𝑘 and 𝑋^𝑡_𝑘 represent the velocity and position arrays of particle 𝑘 at the tth iteration, respectively. 𝑃^𝑡_𝑘 and 𝑃^𝑡_𝑔 represent the local best position for each particle 𝑘 and the global best position among all particles visited so far. 𝑅₁ and 𝑅₂ are 1-by-(𝑚+𝑛−1) arrays in which each digit is 0 or 1. These random arrays are generated from a Bernoulli distribution. O, O^×, and O⁺ are subtraction, multiplication, and addition operators, respectively. The definitions of the operators are as follows.

The subtraction operator (O) defines the differences between the current position of the kth particle, 𝑋^𝑡_𝑘, and a desired position 𝑃^𝑡_𝑘(or 𝑃^𝑡_𝑔). It first finds elements that do not have the same content in 𝑋^𝑡_𝑘 and 𝑃^𝑡_𝑘(or 𝑃^𝑡_𝑔). It schedules those elements based on their orders in 𝑃^𝑡_𝑘(or 𝑃^𝑡_𝑔). Next, it finds elements that have the same content in 𝑋^𝑡_𝑘 and 𝑃^𝑡_𝑘(or 𝑃^𝑡_𝑔) and gives those elements zero values. It puts zero-valued elements in SRD and then schedules them on whatever machine that offers the earliest completion time (ECT). Figure 2 demonstrates the manner in which the O operator works for example 1.

Figure 2 

The subtraction operator (O) applied to example 1.

The multiplication operator (O^×) can enhance our PSO algorithm’s exploration. It first generates a 1-by-(𝑚+𝑛−1) binary vector for a solution vector, and then it does a multiplication process where the asterisk positions from solution are kept. These random binary arrays perform subroutines within PSO that use random numbers to enhance the exploration ability of PSO. Figure 3 demonstrates the manner in which the O^× operator works for example 1. The nonzero-valued elements in 𝐴O^×𝐵 are scheduled first. Then, all zero-valued elements in 𝐴O^×𝐵 are sorted in SRD, and then each zero-valued element is assigned to the machine that offers the ECT.

Figure 3 

The multiplication operator (O×) applied to example 1.

The addition operator (O⁺) is a crossover operator that is commonly used in genetic algorithms. Here, we used a crossover that was proposed in [15]. The crossover scheme has three main steps: (1) it obtains asterisk positions from the first parent 𝐴; (2) it obtains a randomly selected subschedule from the first parent 𝐴; (3) it scans the second parent 𝐵 from left to right and fills the gaps in the child’s (𝐴𝑂⁺𝐵) schedule with jobs taken from the second parent 𝐵. Figure 4 shows an illustration of this crossover scheme.

Figure 4 

The addition operator (O+) applied to example 1.

During the execution of the O, O^×, and O⁺ operators, whenever a complete schedule generates a better solution than the best current solution 𝐶^∗_max, we will update the best current solution 𝐶^∗_max and the global best location (𝑃^𝑡_𝑔).

3.4. Local Search

It is well known that evolutionary memetic algorithms can be improved by hybridization with local search. For each particle 𝑃^𝑡_𝑘, we do the following local search procedure (LSP) to further improve the current solution.

3.4.1. Local Search Procedure (LSP)

Step 1. Initially, set 𝑙=1.

Step 2. Identify machine 𝑖 that has maximum completion time (𝐶_max). Randomly choose one job 𝑗₁ from machine 𝑖 and randomly choose one job 𝑗₂ from machine ℎ(ℎ≠𝑖). Insert job 𝑗₁ after job 𝑗₂ on machine ℎ. If a better 𝐶_max(𝑃^𝑡_𝑘) is found, update 𝑃^𝑡_𝑘 and go to Step 2; otherwise, go to Step 3.

Step 3. Randomly choose two jobs 𝑗₁ and 𝑗₂ from 𝑃^𝑡_𝑘; 𝑗₁ and 𝑗₂ can be on the same machine or on two different machines. Interchange jobs 𝑗₁ and 𝑗₂. If a better 𝐶_max(𝑃^𝑡_𝑘) is found, update 𝑃^𝑡_𝑘 and go to Step 2; otherwise, go to Step 4.

Step 4. If 𝑙=𝐿, stop; otherwise set 𝑙=𝑙+1 and go to Step 2.

Again, during the execution of LSP, whenever a complete schedule generates a better solution than the best current solution 𝐶^∗_max, we will update the best current solution 𝐶^∗_max and the global best location (𝑃^𝑡_𝑔).

3.5. Stopping Criteria and Parameter Settings

We studied the effects of five important parameters (𝑅₁, 𝑅₂, local search moves 𝐿, population size, and number of iterations) on the performance of our proposed PSO. The model was tested and parameterized through a factorial study. The selected PSO parameters were 𝑅₁=0.9, 𝑅₂=0.1, 𝐿=10, population size =73, and number of iterations =4044. The appendix includes a detailed description of our parameterization study.

4. Computational Results

In this section, we present several computational results of the proposed PSO algorithm. We compare our proposed PSO algorithm with a mixed integer programming (MIP) model developed in our previous research [16] on the 𝑅|𝑟_𝑗|𝐶_max problem. The MIP model [16] was coded in AMPL and implemented in CPLEX 11.2. The proposed heuristic, SRD_Reassign, and the proposed PSO algorithm were implemented in C. The MIP model, heuristic, and PSO algorithm were executed on a computer with a 2.5 GHz CPU and 2 GB of memory. Processing times 𝑝_𝑖𝑗 were generated from the uniform distribution [1,100]. Release dates were generated in a manner similar to that of Mönch et al. [17]. We generated release dates 𝑟_𝑗 from the uniform distribution [0,(𝛼/𝑚)((∑^𝑚_𝑖=1∑^𝑛_𝑗=1𝑝_𝑖𝑗)/𝑚)]. 𝛼 controlled the range of release dates. High values of 𝛼 tend to produce widely separated release dates. 𝛼 values were set at 0.1, 0.25, and 0.5. We used 4 machines with 18 jobs (4𝑚18𝑛) to represent small problem instances and 10 machines with 100 jobs (10𝑚100𝑛) to represent large problem instances. For each 𝛼, 20 problem instances were randomly generated. The effectiveness of each algorithm was evaluated by the mean performance and the required computation time (labeled as “Avg. time”). For small problem instances, a ratio was calculated by dividing the algorithm’s makespan by the optimal MIP makespan. For large problem instances, a ratio was calculated by dividing the algorithm’s makespan by the makespan from LB. The mean performance of the algorithm for each 𝛼 was the average ratio obtained from 20 runs of the algorithm.

4.1. Comparison of Heuristics for 𝑅|𝑟_𝑗|𝐶_max Problem

We compared the proposed heuristic SRD_Reassign with the optimal solutions obtained from the MIP model [16] and FCFS. FCFS is a dispatching rule that is commonly used for practical problems with release dates. Computational results for small and large problem instances are given in Tables 2 and 3, respectively. The results show that the proposed SRD_Reassign outperformed FCFS in terms of makespan. For small problem instances, the average SRD_Reassign makespan was 1.08 times the optimum, and the average FCFS makespan was 1.41 times the optimum. Both heuristics outperformed the MIP model in terms of computation time. When 𝛼 was small, both heuristics had larger ratios to the optimal solutions than they had when 𝛼 was large. Also, the MIP took more computation time to find optimal solutions when 𝛼 was small. This probably indicates that problems with small release date ranges are harder to solve than problems with large release date ranges. For large problem instances, the average SRD_Reassign makespan was 1.22-times greater than the lower bound (LB), and the average FCFS makespan was 1.67 times the LB. Both heuristics were calculated very quickly (in less than 1 second) even for large problem instances.

Table 2 

The performance of heuristic SRD_Reassign for small problem instances.

Table 3 

The performance of heuristic SRD_Reassign for large problem instances.

4.2. Comparison of Metaheuristics for 𝑅|𝑟_𝑗|𝐶_max Problem

We compared the proposed PSO with an existing metaheuristic, namely, the version of simulated annealing (SA) described by Lee et al. [18]. This SA variant was originally designed for solving the 𝑃||𝐶_max problem. SA is a metaheuristic, and it can be used without any problem-dependent knowledge; therefore, it can be used to solve the 𝑅|𝑟_𝑗|𝐶_max problem. In order to provide a fair comparison, we used the same initial solution (SRD_Reassign) for both PSO and SA. We also adjusted the SA parameters to ensure that both PSO and SA ran for similar computation times. The termination criterion for SA was set to run for 12 seconds for small problem instances and 83 seconds for large problem instances. If SA found a solution equal to LB, the program would terminate earlier. Computational results for small and large problem instances are given in Tables 4 and 5, respectively.

Table 4 

The performance of PSO for small problem instances.

Table 5 

The performance of PSO for large problem instances.

Computational results show that the proposed PSO outperformed the SA in terms of makespan. For small problem instances, the PSO found optimal solutions at all three 𝛼 settings. The average SA makespan was 1.05 times the optimum. Both metaheuristics outperformed the MIP model in terms of computation time. The last column in Table 4 reports how many times a given algorithm produced a better makespan than the other algorithm. For instance, a value of 𝑐/𝑑 in column PSO/SA means that, out of 20 problems, there were 𝑐 problems for which PSO yielded a better solution than SA, 𝑑 problems for which SA performed better, and 20-𝑐-𝑑 problems for which PSO and SA yielded the same makespan.

For large problem instances, the average PSO makespan was 1.08-times greater than the LB, and the average SA makespan was 1.15 times the LB. The last column in Table 5 shows how many times out of 20 the LB was obtained by 𝐿𝐵₁ and how many times the LB was obtained by 𝐿𝐵₂. When 𝛼 was small, 𝐿𝐵₂ provided a better lower bound than 𝐿𝐵₁; however, when 𝛼 was large, 𝐿𝐵₁ provided a better lower bound than 𝐿𝐵₂. This suggests that 𝐿𝐵₂ performs better for problems with narrow release date ranges, and 𝐿𝐵₁ performs better for problems with wide release date ranges.

4.3. The Effects of the Proposed PSO

Next, since the proposed PSO effectively incorporates a number of ideas (initial solutions, SRD, ECT, and LSP), we examine which parts are essential to its functionality. We examine these effects by disabling a single component, running the proposed PSO without that component, and observing performance. We choose to study large problems. These experiments are described as follows.

PSO-Initial Heuristics: instead of generating an initial population by heuristics, we randomly generated an initial set of solutions to make up the initial population.

PSO-SRD: instead of sorting elements by SRD and then scheduling them on whatever machine that offered the ECT within PSO operators (O and O^×), we randomly chose unscheduled jobs and then scheduled them on whatever machine that offered the ECT.

PSO-ECT: instead of sorting elements by SRD and then scheduling them on whatever machine that offered the ECT within PSO operators (O and O^×), we sorted elements by SRD and then scheduled them on the first available machine.

PSO-LSP: local search procedure was disabled.

𝑃𝑆𝑂-𝐿𝑆𝑃+𝐿𝑆 [13]: local search procedure was disabled and a local search algorithm used in [13] was applied. Since the formulation 0<𝑝_𝑎−𝑝_𝑏<𝐹𝑇_𝑖−𝐹𝑇_𝑗 in step 4 [13] was not suitable for the unrelated parallel machines environment, we modified it to find two jobs from 𝑀_𝑖 and 𝑀_𝑗 such that an exchange of those two jobs was able to improve the current best makespan.

Table 6 lists the average makespan ratio of PSO-variant to standard PSO (which has initial heuristics, SRD, ECT, and LSP by default). Table 6 shows that the PSO performed poorly when the initial heuristics were not applied and when the initial population was randomly generated. The PSO also performed poorly when the ECT strategy was not applied within the PSO operators. The average ratio of PSO without initial heuristics to standard PSO was 1.019, the average of PSO without SRD to standard PSO was 1.006, the average of PSO without ECT to standard PSO was 1.020, and the average of PSO without LSP to standard PSO was 1.007. In all, all of the proposed PSO versions without any one of the parts (heuristic initial solutions, SRD, ECT, and LSP) performed worse than the PSO with all of them.

Table 6 

The effects of the proposed PSO.

Moreover, we compared our proposed PSO with another existing PSO. The closest existing PSO that we were able to find was the hybridized discrete PSO (HDPSO) proposed in [13]. The HDPSO [13] was designed to minimize makespan for identical parallel machines (𝑃||𝐶_max). The proposed PSO and the HDPSO both are designed to minimize makespan for parallel machines and they both use formulas (4)-(5) to update each particle’s velocity and position. However, the HDPSO is still quite different from our PSO. The HDPSO considers problems without release dates; hence, its coding scheme does not consider the order of jobs on the same machine. Also, HDPSO considers an identical parallel machine environment; it uses the LPT rule to assign jobs with zero-valued elements within O, O^×, and O⁺ operators in formulas (4)-(5). It is well known that the LPT rule does not perform well in unrelated parallel machine environments. If HDPSO is used to solve the 𝑅|𝑟_𝑗|𝐶_max problem without modifications, it will not perform very well. Table 7 shows a comparison between HDPSO and our PSO. Since the O, O^×, and O⁺ operators within formulas (4)-(5) are quite different in both PSOs, there is no point in comparing them. We focus our comparison on initial heuristics and local searches. The first column in Table 6 indicates that our PSO with an initial heuristic performs better than a version without an initial heuristic. HDPSO might exhibit similar performance differences. The last column in Table 6 indicates that our proposed LSP performs better than the local search algorithm used in [13]. The standard PSO (LSP is embedded) versus PSO-LSP+LS [13] is 1.000 versus 1.009. Moreover, the proposed LSP outperforms the local search algorithm used in [13] in terms of average computation time. Therefore, we can conclude that our proposed PSO provides better and more efficient strategies for parallel machine makespan minimization problems than what HDPSO provides. Our PSO is more likely to provide promising results than HDPSO.

Table 7 

Comparison between two PSOs.

5. Conclusions and Future Work

We studied the problem of scheduling jobs on unrelated parallel machines with release dates to minimize makespan. In this research, we proposed two lower bounds for the studied problem. We also proposed a heuristic, SRD_Reassign, and a metaheuristic, PSO, to tackle the problem. Computational results showed that SRD_Reassign outperformed the commonly used heuristic, FCFS, in terms of makespan. The proposed PSO outperformed a comparable variant of SA in terms of makespan. Future work can extend our approach for other performance criteria or even for multiobjective parallel machine scheduling problems.

Appendix

We studied the effects of five important parameters (𝑅₁, 𝑅₂, local search moves 𝐿, population size, and number of iterations) on the performance of our proposed PSO. In order to test the significance of each parameter, 10𝑚100𝑛 was chosen as a representative problem instance; the objective was to minimize 𝐶_max in an unrelated parallel machine environment with release dates. Because problems with small ranges of release dates are harder to solve than problems with large ranges of release dates, the release date factor 𝛼 was set to 0.1. In order to obtain information about the importance of each of the factors, we conducted an initial screening experiment. Each parameter was categorized as being at a high or low level, as shown in Table 8. We conducted a 2⁵⁻²_III screening experiment to determine which factors were significant. The results of this experiment are shown in Table 9 where each 𝐶_max value is the average of 20 problem instances. The half normal plot provided by Design Expert indicates that 𝑅₁ (factor 𝐴), population size (factor 𝐷), and iterations (factor 𝐸) were significant to the 𝐶_max response in the screening experiment. The model in terms of coded factors is 𝑦=116.43−1.13𝐴−0.45𝐷−0.59𝐸. The model shows that increasing the 𝐴, 𝐷, and 𝐸 values could decrease makespan.

Table 8 

Factors and levels of PSO parameter study.

Table 9 

Results of 2_III^5–21 factorial design for parameters of PSO.

Next, we used the method of steepest descent (Myers et al. [19]) to provide information about the region of improved response. The path of steepest descent is shown in Table 10. The results show that a reduction in 𝐶_max was experienced after Run 2. Although Run 2 improved 𝐶_max by 2.4% compared with Run 0, its computation time was very long. Run 1 improved 𝐶_max by 2.1% relative to Run 0 and used less computation time. We chose the settings of Run 1 as our final set of parameters. Hence, PSO parameters were set to 𝑅₁=0.9, population size =73, and iterations =4044. Since the 𝑅₂ and 𝐿 were not significant, they were kept at low levels (𝑅₂=0.1 and 𝐿=10) to save computation time.

Table 10 

The path of steepest descent for PSO.