Machine Learning for Dynamic Job Shop Scheduling Problem:

Literature Review

Nawres Boussadia

, Olfa Belkahla Driss

Artificial Intelligence Research Laboratory(LARIA), Higher Business School of Tunis, University of Manouba, Manouba,

Tunisia

Keywords: Dynamic scheduling, Job shop scheduling problem, Machine Learning, Deep learning

Abstract: In the last ten years, Machine Learning (ML) techniques have taken a huge leap forward and researchers have

started to consider ML for job scheduling problems in the industrial field, especially dynamic job shop

scheduling. In this paper, we mainly focus on the dynamic scheduling problem, which is more complex and

difficult to solve and we propose to regroup the methods and approaches used to face it. Therefore, we give a

review of machine and deep learning methods applied to dynamic job shop scheduling problems. In this way,

our work provides a resume of the concerned studies.

1 INTRODUCTION

Among the problems encountered by researchers and

engineers, optimization problems occupy a prominent

place in our time. Formulating optimization problems

and trying to solve them is the main objective of many

researchers. To solve the planning problem, two

objectives must be reconciled. The static aspect

includes the development of implementation plans

based on predictive data. The dynamic aspect is to

make decisions in real time given the state of the

resources and the progress in time of the different

tasks. Workshop scheduling consists of predicting the

sequence of all the elementary operations required to

carry out manufacturing orders on production

resources while taking into account internal and

external constraints. In a complex production

environment, scheduling can become an extremely

difficult problem to solve. Production scheduling is

the determination of the order in which a number of

jobs are to be executed. This determination concerns

the planning of the use of available human and

machine resources in order to better control the costs

and to master the manufacturing delays of the decided

productions. The resolution of a scheduling problem

goes through an identification and modelling phase

and a research phase of the adequate resolution

method. The development of new technologies in the

https://orcid.org/0000-0002-0274-0702

https://orcid.org/0000-0003-3077-6240

manufacturing world has allowed manufacturers to

meet the ever-increasing challenge of dealing with

multiple objectives and unforeseen events to which

they are subjected, such as a change or cancellation

of a production order, or the arrival of a rush order.

Most planning problems are, or come down to,

dynamic optimization problems.

In our work, we pay more attention to a dynamic

job-shop scheduling problems in order to group and

present the different optimization methods developed

in the literature and the different criteria to optimize.

The remaining sections of the paper are organized

as follows: In Section 2, we defined the dynamic job

scheduling problem in detail. Section 3 is devoted to

a comprehensive review of the literature on dynamic

shop scheduling. A complete review of the literature

on dynamic shop scheduling with machine learning is

illustrated in Section 4. The discussion of the various

contributions is presented in section 5. We end this

paper with a general conclusion that summarizes the

different phases of work in section 6.

2 DESCRIPTION OF DYNAMIC

JOB SCHEDULING PROBLEM

A collection of jobs to be done on a set of resources

in order to optimize the objective function describes

444

Boussadia, N. and Belkahla Driss, O.

Machine Learning for Dynamic Job Shop Scheduling Problem: Literature Review.

DOI: 10.5220/0010736200003101

In Proceedings of the 2nd International Conference on Big Data, Modelling and Machine Learning (BML 2021), pages 444-450

ISBN: 978-989-758-559-3

the scheduling problem. The scheduling problem

resides in the continuous adaptation of the process of

a set of resources to execute a set of tasks to the real

situation of the considered system. This type of

problem is often encountered in manufacturing

workshops that work to order where the delivery time

represents one of the major difficulties. The job shop

problem is one of the most studied and most difficult

problems in scheduling theory (NP-hard). It is

essential to know, in the resolution of this type of

problem, if one must privilege the quality of the

sought solution, the speed of the calculation time or

find a compromise. The optimal solution is therefore

impossible in most cases because of its combinatorial

character. Therefore, we generally resort to the so-

called approximate methods, which give approximate

solutions in a reasonable time.

Dynamic scheduling problems formulated as

optimization problems are often classified as NP-

hard, especially those related to production systems.

The solution of such problems requires dedicated

methods; while exact methods cannot solve this type

of problems due to the huge computational time,

approximate methods offer the possibility to find a

feasible solution in a reasonable time.

In our paper, we are particularly interested in real-

time scheduling problems for a job shop system in a

dynamic environment. This system is subject to a

disturbance of the environment represented by the

occurrence of new urgent orders, which it must

execute, machine breakdowns, unexpected

processing delays and cancellation of orders. The

problem imposed here is how the system should react

to the occurrence of one of these events.

3 SCHEDULING APPROACHES

ON DYNAMIC JOB SHOP

SCHEDULING PROBLEM

Production scheduling systems are generally applied

in real-world manufacturing systems under the

impact of unpredictable or dynamic events including

machine breakdowns, unforeseen processing delays,

random arrivals of urgent orders, and order

cancellations. Because of these dynamic occurrences,

the initial scheduling strategy is suboptimal and/or

infeasible. Therefore, a proper dynamic rescheduling

approach is needed to deal with these dynamic events.

To cope with a dynamic job shop scheduling

problem that takes into consideration random work

arrivals and machine failures, (Zandieh and Adibi,

2010) proposed a variable neighbourhood search

(VNS) based scheduling approach. They selected an

event-based policy to deal with the problem's

dynamic nature. An artificial neural network with an

error backpropagation learning algorithm is used to

adjust the VNS parameters at every rescheduling

point based on the issue circumstances to increase the

efficiency of the scheduling technique.

The dynamic flexible job shop scheduling

problem (DFJSSP) with publication dates was

explored by (Nie et al., 2013). For the dynamic

scheduling problem, they present a heuristic for

implementing reactive scheduling. They also suggest

using Genetic Expression Programming (GEP) to

build reactive scheduling strategies for dynamic

scheduling automatically. Three factors, such as shop

floor utilization, proximity to due date, and problem

flexibility, are considered in the simulation

experiments in order to evaluate the performance of

the reactive scheduling policies constructed by the

proposed genetic expression programming-based

approach under a variety of processing conditions.

The simulation considers the minimization of

makespan, average flow time, and average latency as

scheduling performance measures.

(Kundakc and Kulak, 2016) presented a dynamic

shop floor scheduling issue in which new tasks arrive,

a machine fails, and the processing time changes.

Heuristic techniques are effective for tackling

dynamic shop floor scheduling issues since they are

NP-hard combinatorial optimization problems. The

authors offer hybrid GA (Genetic Algorithms)

methods in which a novel KK + exchange heuristic

and well-known dispatching rules (SPT, LPT, SRPT,

and LRPT) + exchange are combined with a GA

algorithm to give quick and efficient solutions to

large dynamic shop scheduling issues.

In order to solve the dynamic scheduling

problems of flexible job shop, the authors (Ning et al.,

2016) proposed an improved multiphase hybrid

quantum particle swarming algorithm. First, they

planned a dual-chain structure coding method

including a machine distribution chain and a process

chain. Then, they proposed a dynamic periodic and

event-driven scheduling strategy. Finally, these

authors applied a new method to the

Brandimarte(1993) set, including 10 examples, used

for the verification of the effectiveness of the

proposed method and for dynamic simulation.

(Wang et al., 2017) developed a dynamic

rescheduling approach based on a variable interval

rescheduling strategy (VIRS) to cope with a job

shop's flexible dynamic scheduling problem by

considering machine failures, arrival of urgent work,

and work damage as disruptions. They suggest, on the

Machine Learning for Dynamic Job Shop Scheduling Problem: Literature Review

445

other hand, an enhanced genetic algorithm (GA) to

reduce the makespan. A random initialization

population mixture is meant to create a high-quality

starting population in our enhanced GA by combining

an initialization machine and an initialization

operation with random initialization. To prevent

slipping into the local optimal solution, the elitist

strategy (ES) and the enhanced population diversity

strategy (IPDS) are utilized.

Besides the classical scheduling approaches, such

as metaheursitics, the recent studies are focused on

machine/deep learning. In the following section, we

present the recent studies based on machine/deep

learning for dynamic scheduling problems.

4 MACHINE LEARNING BASED

METHODS ON DYNAMIC JOB

SHOP SCHEDULING

PROBLEM

The scheduling of production processes has long been

the subject of extensive research. Dynamic job shop

scheduling problems are known in the literature as the

most difficult problems to solve.

The difficulty lies in the choice of the best

approach for their solution as well as in the

determination of the best scheduling in reasonable

times, as close as possible to the optimal solution.

Several authors have studied the solution of these

types of scheduling problems using Machine

Learning. The authors (Bouazza et al., 2017) used

intelligent products (IPs) as a solution to solve shop

floor scheduling problems. The product gathers

information from the existing planning environment

and needs machines to perform a set of tasks. The IP

divides the choice into two steps to schedule

production tasks in a simulated environment: choose

a machine and reschedule the selected queue. The

following scenario and configurations were used to

test the proposed model: Six different SPs (service

providers) compose the manufacturing cell, and nine

families are estimated to add sufficient complexity to

the scheduling challenge. They used the Boltzmann

distribution rule by Q-learning method to create

stochastic decisions. The PIs gradually converge to an

ideal behavior as a result of this two-step decision

process. We use the most popular machine selection

criterion, shortest tail, to compare the Q-learning

method. It is used in conjunction with four different

allocation rules: First In, First Out, Shortest Job First,

Highest Priority First, and Last in First Out. With

respect to Cmax, the results are almost identical.

By taking machine failures into account, the

authors (Zhao et al., 2019) suggested an enhanced Q-

learning method with dual-layer actions to address the

dynamic flexible job scheduling problem (DFJSP).

The suggested Q-learning Agent based on

dispatching rules achieves the initial scheduling

scheme, while the Genetic Algorithm (GA) obtains

the rescheduling approach. Experiments based on the

FJSP issue Mk03 are created and executed to show

this method. The findings show that, as compared to

utilizing a single SPT, FIFO, or EDD method, the

agent-based system can choose the optimum strategy

for diverse machine failures, increasing efficiency.

This demonstrates the effectiveness of the suggested

Q-learning in a dynamic and flexible job-shop setting.

A deep Q network (DQN) was developed by

(Luo, 2020) to deal with the ongoing production

status and learn the most suitable action (i.e., dispatch

rule) at each rescheduling point. When an operation

is done or a new task comes, he suggested six

compound dispatch rules to pick an operation at the

same time and allocate it to operable machines. As a

result, seven general state characteristics are derived

to characterize the rescheduling points' production

status. The state action value (Q-value) of each

dispatch rule may be determined by feeding the

continuous state characteristic into DQN. Deep Q-

learning (DQL) is used to train the proposed DQN,

which is further enhanced by two enhancements: dual

DQN and smooth update of target weights.

Furthermore, in the practical implementation of the

trained DQN, a "softmax" action selection policy is

employed to favour rules with higher Q-values while

retaining the policy's entropy. In a simulated flexible

shop with 30 machines, 20 beginning tasks, and 100

additional insertions, the suggested DQN is trained.

They put the DQN and other comparison criteria to

the test on 20 distinct cases, each with 20 independent

replications. DQN outperforms the suggested

composite dispatch rules and other well-known

dispatch rules in both trained and untrained

production configurations, according to the findings.

Simultaneously, comparisons of the DQN and the Q-

stand learning agent demonstrated the DQN's

superiority in handling the continuous state space.

(Liu et al., 2020) views JSSP to be a sequential

decision-making issue and suggests that it be solved

via deep reinforcement learning. An actor network

and a critical network, both of which comprise

convolution layers and a fully connected layer, are

included in the proposed model. The agent in the

actor network learns how to act in various scenarios,

while the critic network assists the agent in evaluating

the statement's worth before returning to the actor

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

446

network. To train the model, this paper offers a

parallel learning approach that combines

asynchronous updating and a deep deterministic

policy gradient (DDPG). Different basic dispatching

rules are considered actions, and the entire network is

trained in parallel in a multi-agent environment. The

evaluation is based on more than ten examples from

the OR-Collection, a well-known reference problem

library. The results show that the model can cope with

unexpected incidents, such as a machine failure or a

sudden extra order. Furthermore, the quality of the

solutions found by is also comparative, outperforms

traditional dispatch rules, and runs almost as fast as

simple dispatch rules.

To cumulate the advantages of real-time response

and flexibility of deep convolutional neural networks

(CNNs) and reinforcement learning (RL), the authors

(Han and Yang, 2020) proposed a deep reinforcement

learning (DRL) framework, to acquire behavioral

strategies from the captured manufacturing state.

Moreover, it is more suitable for real-world

manufacturing problems related to control. In this

framework, the scheduling process using disjunction

graphs is considered as a multi-step sequential

decision problem, and deep CNN is used to

approximate the value of the execution state. The

execution state is represented as a multi-channel

image and input to the network. Static computation

experiments were performed on 85 instances of JSSP

(Job Shop Scheduling Problem) in the well-known

OR library. The results indicate that the proposed

algorithm can obtain optimal solutions for small-scale

problems, and outperforms any single heuristic rule

for large-scale problems, with performance

comparable to genetic algorithms.

The authors offer a new dynamic scheduling

technique based on Petri nets via DQN with graph

convolutional network (Hu et al., 2020). (GCN).

First, based on operation sequence, resource usage

restrictions, and processing time, flexible

manufacturing systems (FMSs) are modeled using

timed S

PR (system of simple sequential processes

with resources). Then, a PNC layer with two graph

convolution sublayers is built to implement feature

propagation from a location to a transition and a

transition to a location, respectively, according to the

unique sub-network structure of the S

PR time. The

advantage of the PNC (Petri-net convolution) layer is

that the number of its training parameters is only

related to the number of filter channels, and has

nothing to do with the scale of the S

PR clock.

Therefore, it is possible to overcome the problem of

parameter explosion when building a deep neural

network. In comparison to heuristic approaches and a

DQN with basic multilayer perceptrons, the

experimental findings demonstrate that the proposed

DQN with a PNC network may give superior answers

to dynamic planning issues in terms of manufacturing

performance, computing efficiency, and flexibility.

For online scheduling in flexible manufacturing

systems, the author (Bar et al., 2020) proposes a

reinforcement learning (RL) technique (FMS).

Scheduling becomes more complicated when

numerous activities with various optimization goals

are considered, and unexpected occurrences result in

prolonged downtime until a new plan is formed. As a

solution to this problem, they use the MARL (Multi-

Agent RL) version of Deep Q-Networks (DQN)

without explicit information exchange between

agents, with agents that have learned to efficiently

guide products through the factory and achieve near-

optimal timing regarding resource allocation. These

agents control the products and can react to

unexpected machine failures and reconfigurations of

the module (machine) topology. An FMS with six

different manufacturing modules and six positions in

which they can be placed is used. All experiments

start with a fixed number of three agents controlling

three identical products in the order stack in the initial

state of each period. The agents must complete a fixed

job specification with four operations, each operation

having two alternative manufacturing modules with

different processing time required to execute the

operations.

(Kardos et al., 2021) applied a reinforcement

learning method, including Q Learning, to reduce the

average lead-time of production orders in the shop

floor production system. The intelligent product

agents (intelligent elements, called agents, are able to

learn the best decision strategy during a learning

process in the RL to maximize the overall

optimization objective) can select a machine for each

production step based on real-time information. The

performance of the strategies learned by the RL agent

was tested against standard heuristics. Since these

heuristics are known to be applicable to simple

scenarios, it is important to understand how

complicated the implementation of RL has become

from a practical standpoint. To this end, by increasing

the total number of process steps used and reducing

the defined time interval between dynamically

generated PIs, a set of problems with increased

complexity was defined. In the addition, the

performance of the policy is tested when the agent has

the ability to directly select an IP, compared to the

case where the agent is limited to choosing between

standard dispatch rules. The simulation model is

infused from Bouazza et al. (2017), however in this

Machine Learning for Dynamic Job Shop Scheduling Problem: Literature Review

447

case the production process may have multiple

phases, the distribution of manufacturing orders in

each concentration is random, and only one decision

needs to be made, namely the choice of machine. The

simulation model and decision logic are implemented

in the Python programming language and use the

DES framework based on the SimPy process.

Another important aspect of efficiency is the tight

integration of the decision logic, which has been

implemented using the Tensorflow library.

In the end of this section, we note that Machine

Learning is also applied to Flow shop scheduling

problems, such as (Zhang et al., 2013) and (Xue et al.,

2018). A summary of related research is presented in

Table 1.

Table 1: Summary of relevant studies by machine learning

algorithms

Refere

nce

Approa

Optimization

criteria

Dynamic events

Bouazz

a et al.,

2017

Q-

learnin

g +

SMA

Average

Waiting

Times &

Weighted

Average

Waiting

Time

frequent arrivals of

work, variation in

processing times and

set-up times

Zhao et

al.,

2019

Q-

learnin

+ GA

Time of

delay

machine breakdowns

Luo,

2020

DQL Total

tardiness

new job insertions

Liu et

al.,

2020

DL +

RL +

SMA

Makespan a machine breakdown

+ a sudden additional

orde

Han

and

Yang,

2020

CNN +

Makespan order-driven

manufacturing

Hu et

al.,

2020

Petri

net via

DQN

with

GCN

Makespan shared resources, route

flexibility and

stochastic raw product

arrivals

Bar et

al.,

2020

RL +

SMA

Makespan unexpected machine

failures + plant

reconfigurations +

module (machine)

topology

reconfi

urations

Kardos

et al.,

2021

Q-

learnin

g + RL

Average lead

time

Fluctuating customer

demands, expected

short delivery times

and the need to

confirm orders quickl

Machine learning is not widely applied to the

static case, we can cite (Wang and Usher, 2004)

(Wang and Usher, 2005) (Lin et al., 2019) (Hameed

and Schwung, 2020).

5 DISCUSSION

Scheduling is a field of operations research and

production management concerned with increasing a

company's efficiency in terms of production costs and

delivery timeframes. From manufacturing to

information technology, all economic sectors have

scheduling issues. Operational planning for industrial

production systems involves managing the allocation

of resources over time while optimizing a set of

standards. It also means scheduling the execution of

a project by allocating resources to tasks and setting

their execution dates. The problems of resource

allocation, task organization, meeting deadlines and

making decisions in a timely manner are all

difficulties that must be overcome in the management

of production systems in an industrial environment.

Due of their limited real-time responsiveness,

traditional methods to job shop scheduling difficulties

are currently unsuitable for complex and dynamic

production settings. Naturally, these methods must

have a solid theoretical foundation to approach these

complex problems with some confidence in the

quality of the result. Among the most difficult

scheduling problems are those related to dynamic job

shop. Their optimal solution is, in most cases, very

difficult because of their combinatorial character. In

this paper, we have summarized the studies that focus

on this problem and studied the application of

machine learning (RL, DRL, DQL, CNN...) to a

workshop production scheduling process considering

dynamic events such as random job arrivals, machine

breakdowns, job disturbances and changes in

processing time with a criterion optimization

objective.

6 CONCLUSION AND FUTURE

RESEARCH LINES

This conclusion explains our approach in this article.

Our goal is to be exhaustive and to make the reader

aware of problems whose importance escapes less

and less computer scientists, but still too many

industrialists. Therefore, this article is structured as

follows. First, we presented the static job shop

problem by detailing the approaches used in each of

the cited articles, the criterion (s) to be optimized, the

types of scheduling problems solved, and the

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

448

benchmark used. Secondly, we have done the same

with the dynamic job shop and flow shop problems

by taking into account dynamic events.

Therefore, in this paper we are interested in the

consolidation of the literature review of the static job

shop scheduling problem without machine learning,

as well as the job shop and flow shop problems with

machine learning in real time (dynamic).

Several dynamic scheduling methods have been

presented, including, on the one hand, heuristics,

meta-heuristics, and multi-agent systems, and on the

other hand, machine/deep learning algorithms such as

Q-learning, reinforcement learning, Deep

reinforcement learning, and Deep Q-learning.

Although there has been some research on dynamic

scheduling systems, more effort is still needed to deal

with these NP-hard scheduling problems. In the

future, more aspects can be considered to expand the

research by adopting new optimization methods, such

as Biogeography-based optimization, to solve the

dynamic JSSP, and by adopting new approaches

based on deep learning, such as Convolutional neural

networks.

REFERENCES

Bar, S., Turner, D., Mohanty, P.K., Samsonov, V.,

Bakakeu, J.R., Meisen, T., 2020. Multi Agent Deep Q-

Network Approach for Online Job Shop Scheduling in

Flexible Manufacturing 10.

Bouazza, W., Sallez, Y., Beldjilali, B., 2017. A distributed

approach solving partially flexible job-shop scheduling

problem with a Q-learning effect. IFAC-Pap. 50,

15890–15895.

Hameed, M.S.A., Schwung, A., 2020. Reinforcement

Learning on Job Shop Scheduling Problems Using

Graph Networks.

https://doi.org/10.13140/RG.2.2.13862.96326.

Han, B.-A., Yang, J.-J., 2020. Research on Adaptive Job

Shop Scheduling Problems Based on Dueling Double

DQN. IEEE Access 8, 186474–186495.

Hu, L., Liu, Z., Hu, W., Wang, Y., Tan, J., Wu, F., 2020.

Petri-net-based dynamic scheduling of flexible

manufacturing system via deep reinforcement learning

with graph convolutional network. J. Manuf. Syst. 55,

1–14.

Kardos, C., Laflamme, C., Gallina, V., Sihn, W., 2021.

Dynamic scheduling in a job-shop production system

with reinforcement learning. Procedia CIRP 97, 104–

109.

Kundakcı, N., Kulak, O., 2016. Hybrid genetic algorithms

for minimizing makespan in dynamic job shop

scheduling problem. Comput. Ind. Eng. 96, 31–51.

Kundakcı, N., Kulak, O., 2016. Hybrid genetic algorithms

for minimizing makespan in dynamic job shop

scheduling problem. Comput. Ind. Eng. 96, 31–51.

Lin, C.-C., Deng, D.-J., Chih, Y.-L., Chiu, H.-T., 2019.

Smart Manufacturing Scheduling With Edge

Computing Using Multiclass Deep Q Network. IEEE

Trans. Ind. Inform. 15, 4276–4284.

Liu, C.-L., Chang, C.-C., Tseng, C.-J., 2020. Actor-Critic

Deep Reinforcement Learning for Solving Job Shop

Scheduling Problems. IEEE Access 8, 71752–71762.

Luo, S., 2020. Dynamic scheduling for flexible job shop

with new job insertions by deep reinforcement learning.

Appl. Soft Comput. 91, 106208.

Nie, L., Gao, L., Li, P., Li, X., 2013. A GEP-based reactive

scheduling policies constructing approach for dynamic

flexible job shop scheduling problem with job release

dates. J. Intell. Manuf. 24, 763–774.

Nie, L., Gao, L., Li, P., Li, X., 2013. A GEP-based reactive

scheduling policies constructing approach for dynamic

flexible job shop scheduling problem with job release

dates. J. Intell. Manuf. 24, 763–774.

Ning, T., Huang, M., Liang, X., Jin, H., 2016. A novel

dynamic scheduling strategy for solving flexible job-

shop problems. J. Ambient Intell. Humaniz. Comput. 7,

721–729.

Ning, T., Huang, M., Liang, X., Jin, H., 2016. A novel

dynamic scheduling strategy for solving flexible job-

shop problems. J. Ambient Intell. Humaniz. Comput. 7,

721–729.

Wang, L., Luo, C., Cai, J., 2017. A Variable Interval

Rescheduling Strategy for Dynamic Flexible Job Shop

Scheduling Problem by Improved Genetic Algorithm.

J. Adv. Transp. 2017, 1–12.

Wang, L., Luo, C., Cai, J., 2017. A Variable Interval

Rescheduling Strategy for Dynamic Flexible Job Shop

Scheduling Problem by Improved Genetic Algorithm.

J. Adv. Transp. 2017, 1–12.

Wang, Y.-C., Usher, J.M., 2004. Learning policies for

single machine job dispatching. Robot. Comput.-Integr.

Manuf. 20, 553–562.

Wang, Y.-C., Usher, J.M., 2005. Application of

reinforcement learning for agent-based production

scheduling. Eng. Appl. Artif. Intell. 18, 73–82.

Xue, T., Zeng, P., Yu, H., 2018. A reinforcement learning

method for multi-AGV scheduling in manufacturing,

in: 2018 IEEE International Conference on Industrial

Technology (ICIT). Presented at the 2018 IEEE

International Conference on Industrial Technology

(ICIT), IEEE, Lyon, pp. 1557–1561.

Xue, T., Zeng, P., Yu, H., 2018. A reinforcement learning

method for multi-AGV scheduling in manufacturing,

in: 2018 IEEE International Conference on Industrial

Technology (ICIT). Presented at the 2018 IEEE

International Conference on Industrial Technology

(ICIT), IEEE, Lyon, pp. 1557–1561.

Zandieh, M., Adibi, M.A., 2010. Dynamic job shop

scheduling using variable neighbourhood search. Int. J.

Prod. Res. 48, 2449–2458.

Zandieh, M., Adibi, M.A., 2010. Dynamic job shop

scheduling using variable neighbourhood search. Int. J.

Prod. Res. 48, 2449–2458.

Machine Learning for Dynamic Job Shop Scheduling Problem: Literature Review

449

Zhang, Z., Wang, W., Zhong, S., Hu, K., 2013. FLOW

SHOP SCHEDULING WITH REINFORCEMENT

LEARNING. Asia-Pac. J. Oper. Res. 30, 1350014.

Zhang, Z., Wang, W., Zhong, S., Hu, K., 2013. Flow shop

scheduling with rienforcement learning. Asia-Pac. J.

Oper. Res. 30, 1350014.

Zhao, M., Li, X., Gao, L., Wang, L., Xiao, M., 2019. An

improved Q-learning based rescheduling method for

flexible job-shops with machine failures, in: 2019 IEEE

15th International Conference on Automation Science

and Engineering (CASE), Canada, pp. 331–337.

BML 2021 - INTERNATIONAL CONFERENCE ON BIG DATA, MODELLING AND MACHINE LEARNING (BML’21)

450