A hybrid method of 2-TSP and novel learning-based GA for job sequencing and tool switching problem

作者：Ahmadi Ehsan^*; Goldengorin Boris; Suer Gursel A; Mosadegh Hadi

来源：Applied Soft Computing, 2018, 65: 214-229.

DOI：10.1016/j.asoc.2017.12.045

• Combinatorial optimization
• Job scheduling
• Tool switches
• Genetic algorithm
• Q-learning
• Reinforcement learning

One of the well-known problems in single machine scheduling context is the Job Sequencing and Tool Switching Problem (SSP). The SSP is optimally sequencing a finite set of jobs and loading restricted subset of tools to a magazine with the aim of minimizing the total number of tool switches. It has been proved in the literature that the SSP can be reduced to the Job Sequencing Problem (JSeP). In the JSeP, the number of tool switches from the currently processed job to the next job depends on the sequencing of all predecessors. In this paper, the JSeP is modeled as a Traveling Salesman Problem of Second Order (2-TSP). We call the induced JSeP by 2-TSP as the Job Sequencing Problem of Second Order (2-JSeP) with a different objective function formulation from JSeP and prove that 2-JSeP is NP-hard. Then the Assignment Problem of Second Order (2-AP) and Karp-Steele patching heuristic are incorporated to solve 2-JSeP. The obtained solution, however, does not guarantee the optimal sequence and are used to seed a Dynamic Q-learning-based Genetic Algorithm (DQGA) to improve the solution quality. Q-learning, which is a kind of reinforcement learning method, is used to learn from the experience of selecting the order of mutation and crossover operators in each generation of the genetic algorithm. The computational results on 320 benchmark instances show that the proposed DQGA is comparable to the state-of-the-art methods in the literature. The DQGA even outperforms the existing methods for some instances, as could improve the reported best-known solutions in notably less time. Finally, through the statistical analysis, the performance of DQGA is compared with those of non-learning genetic algorithms.

• 出版日期2018-4