Many-Objective Genetic Programming for Job-Shop Scheduling

This article overviews a recent PhD dissertation representing the first effort at many-objective optimization in job shop scheduling (JSS). The thesis develops genetic programming hyperheuristic (GP-HH) approaches to evolve effective dispatching rules for many conflicting objectives in JSS problems. The aim is to develop GP-HH methods that alleviate issues related to many-objective optimization in JSS problems and evolve new effective dispatching rules capable of enhancing job shops' productivity.


Introduction
This article overviews a recent PhD dissertation representing the first effort at manyobjective optimization in job shop scheduling (JSS). The thesis develops genetic programming hyperheuristic (GP-HH) approaches to evolve effective dispatching rules for many conflicting objectives in JSS problems. The aim is to develop GP-HH methods that alleviate issues related to many-objective optimization in JSS problems and evolve new effective dispatching rules capable of enhancing job shops' productivity.

Background
Job shop scheduling (JSS) [7] is a non-deterministic polynomial-time (NP) hard combinatorial optimization problem that involves assigning different manufacturing jobs to machines at specific times while trying to minimize a number of objectives, including the mean flowtime (mF), maximal flowtime (maxF), mean weighted tardiness (mWT) and maximal weighted tardiness (maxWT). These objectives are also considered to minimize in this thesis. JSS problems have drawn a lot of interest from academics and industry experts. As reported by Johns and Rabelo [5], thousands of manufacturers contribute billions of dollars to the United States' economy. Furthermore, JSS is considered one of the significant production scheduling problems in practice. It has a wide range of applications in many industries such as cloud computing [8] and management and operations research [6]. JSS has received substantial research attention due to its high computational challenges and strong practical value. A JSS problem usually has a set of machines on the shop floor that can be used to process a set of jobs [7]. Each job has a predetermined sequence of operations, which needs to be carried out to complete the job. An example of a job shop studied in this thesis is shown in  Dispatching rules have been applied extensively to JSS problems due to their computational efficiency. Dispatching rules can be seen as a priority function which is used to assign priority to each job waiting to be processed by a machine. Then, the next job to process will be selected based on the priority value. Such computation is carried out at each decision point (e.g., when a machine becomes idle) and can be done efficiently. This can be seen in  Genetic Programming (GP) has been a promising approach for designing dispatching rule heuristics automatically because GP has an ability to evolve priority functions with its flexible representation. The hyper-heuristic approach that uses GP to solve JSS problems is known as GP based Hyper-Heuristic (GP-HH ) [1]. This thesis evolves dispatching rules for both static and dynamic JSS problems automatically.
Many-objective optimization has become an active research topic [2]. As emphasized by Deb in [2], a large proportion of real-world problems can be described naturally as manyobjective problems (MaOPs). A class of optimization problems with more than three objectives is referred to as many-objective optimization. The last decade has witnessed the emergence of many-objective optimization as a booming topic in a wide range of complex modern real-world scenarios.
Scheduling theory has been established over the years. Still, most existing literature on the automatic design of dispatching rules mainly concentrates on single-objective JSS [3] and multi-objective JSS [4].
Only a few algorithms in the literature tackle JSS problems with more than three scheduling objectives. These algorithms utilize the conventional MOEAs (NSGA-II and SPEA2), but MOEAs experienced substantial difficulties when they were adopted to tackle MaOPs [2]. It is both theoretically and practically important to develop innovative GP-HH algorithms for many-objective JSS.

Objectives
This research was broken down into the following key objectives: 1. Investigate how GP can be used to handle many-objective JSS problems. 2. Investigate how to develop GP-HH approaches for the non-uniform Pareto front of many-objective JSS problems which can evolve high-quality Pareto-optimal dispatching rules. 3. Investigate how to hybridize a local search with a global search and improve the quality of the evolved rules in many-objective JSS.

Contributions
The main contributions of this thesis are to propose: 1. The first many-objective GP-HH method for JSS problems to find the Pareto-fronts of non-dominated dispatching rules. In order to visualize the interdependencies between different objectives, this thesis used the aggregated Pareto-front. Part of this contribution has been published at: