Autonomous control methods in job shop logistics - Methoden voor autonome controle in job shop logistiek

(1)

Delft University of Technology

FACULTY OF MECHANICAL, MARITIME AND MATERIALS ENGINEERING

Department of Marine and Transport Technology Mekelweg 2 2628 CD Delft the Netherlands Phone +31 (0)15-2782889 Fax +31 (0)15-2781397 www.mtt.tudelft.nl

Specialization: Transport Engineering and Logistics

Report number: 2017.TEL.8182

Title:

Autonomous control methods in job

shop logistics

Author:

R.O. de Boer

Title (in Dutch) Methoden voor autonome controle in job shop logistiek

Assignment: Literature

Confidential: No

Supervisor: Dr.ir. Y. Pang

(2)

Delft University of Technology

FACULTY OF MECHANICAL, MARITIME AND MATERIALS ENGINEERING

Department of Marine and Transport Technology Mekelweg 2 2628 CD Delft the Netherlands Phone +31 (0)15-2782889 Fax +31 (0)15-2781397 www.mtt.tudelft.nl

Student: R.O. de Boer Assignment type: Literature

Supervisor (TUD): Dr.ir. Y. Pang (TU Delft) Creditpoints (EC): 10 Specialization: TEL

Report number: 2017.TEL.8182 Confidential: No

Subject: Autonomous control methods in job shop logistics

The logistics in job shop production is becoming more and more complex with increasing demands from stakeholders. The improvement of the logistic processes is important for job shops to be competitive. Along with the classical ways of conventional logistic control and problem solving, autonomous control can greatly increase the flexibility and efficiency in jobs shop manufacturing.

This assignment is to provide an overview of autonomous control methods used to solve the logistic problems in job shop production systems. The state of the art of the methods and applications will be surveyed based on academic resources and industrial practices. The main tasks of this literature assignment should cover the following:

• To review the concepts and development of autonomous control in logistics;

• To investigate diverse autonomous control methods applied for job shop production systems; • To categorize the control methods used to solve the logistic problems in job shops;

• To summarize the impacts of autonomous control on job shop logistics.

This report should be arranged in such a way that all data is structurally presented in graphs, tables, and lists with belonging descriptions and explanations in text.

The report should comply with the guidelines of the section. Details can be found on the website. The mentor,

(3)

Summary

Recently a lot of research has been done on autonomous control as a replacement for centralised control. This report focuses on the implementation of autonomous control in job shops.

Autonomous control in logistics is: “A set of logistic objects or entities that are structured in a decentralized heterarchical system which can render their own decisions and to take actions to control something”. The entity that controls the system is not central anymore but has been implemented in logistical objects that each control their own decisions. All the objects are operating in a heterarchy which mean they are all equal. Recent developments in ICT has enabled these logistic objects to interact in such way. Autonomous control will let the system organise itself to be successful, but it will also make the system unpredictable, it is not clear how the system will behave upfront.

Autonomous control can greatly improve the logistical processes in a job shop. In a job shop there are different machines which each can execute different operations. This type of factory is used for production of small batches of highly customized products. Planning which machine does what kind of operation and on which part is quite complex, this is called the job shop scheduling problem.

To solve the complex scheduling problems, several autonomous control methods have been developed. Some have been specifically designed for job shop control while others are distilled from other practices. A lot of these distilled methods come from nature and then specifically from ant’s pheromone. Th following list summarizes the different methods.

 Methods specifically designed for autonomous control in job shops o Queue length estimator, searches for the shortest queue

o Due date method, the process with the latest due date goes first

o Simple rule based 2, previous data is compared to check which option is the best o One logistic target per rule, machines send a attraction signal based on specified rules o Ring like topology, a machine chooses a task from a table and then moves the table to the

next machine

 Methods from other practises

o Holonic manufacturing, machines place bids on orders managed by a management agent o Market based control, parts place bids on machines

o Artificial potential fields, parts are attracted and repelled by artificial fields

o Distributed logistics routing protocol; parts and machines make their own plan and try to solve their differences

(4)

o Link-state internet routing protocol, a map of links between machines is plotted and the parts choose the shortest path

 Methods inspired by nature

o Bee foraging, based on how bees search for honey o Bacterial chemotaxis, based on how bacteria move

o Bionic manufacturing system, the system tries to solve local and global problems by attraction fields

 Methods inspired by ants pheromone, these methods are all based on how ants forage for food by using pheromones.

o Pheromone method

o Pheromone based coordination o Ant colony control

With so many methods for autonomous control in job shops, it is not clear which method works best for what kind of setup in the job shop. Also, it is important to note that autonomous control is not always better than traditional problem solving techniques. Combining different autonomous control methods can lead to better but also to worse results compared to a single method. One more thing to note is that implementing autonomous control methods can cause dynamical behaviour, if done well this periodical behaviour can be used to improve the system but it can also lead to dangerous results.

(5)

Table of figures

FIGURE 1 FUTURE CONDITIONS AND REQUIREMENTS ON LOGISTIC PROCESSES [3] 5 FIGURE 2 CATALOGUE OF CRITERIA FOR ANALYSING THE LEVEL OF AUTONOMY IN A LOGISTICS SYSTEM [5] 8 FIGURE 3 EVALUATION SYSTEM FOR AUTONOMOUS CONTROL METHODS IN COUPLED PLANNING AND CONTROL

SYSTEMS [10] 9

FIGURE 4 A TYPICAL JOB SHOP FLOOR [12] 11

FIGURE 5 MATRIX MODEL OF A SHOP FLOOR [16] 15

FIGURE 6 TRADITIONAL TECHNIQUES [17] 16

FIGURE 7 ANALYTICAL TECHNIQUES [17] 17

FIGURE 8 HEURISTIC TECHNIQUES [17] 18

FIGURE 9 OVERVIEW OF THE DIFFERENT CATEGORIES OF AUTONOMOUS CONTROL METHODS “OWN WORK” 19

FIGURE 10 RING ARRANGEMENT OF THE HOLONS [22] 21

FIGURE 11 UML SEQUENCE DIAGRAM FOR THE PROPOSED MODEL [22] 23

FIGURE 12 SYSTEM CONFIGURATION [24] 24

FIGURE 13 INTERACTION BETWEEN ENTITIES IN 2D [26] 25

FIGURE 14 EVOLUTION OF THE ATTRACTIVENESS OF 𝑅𝑗 OVER TIME [26] 28

FIGURE 15 AN EXAMPLE OF A DISTANCE CALCULATION [26] 29

FIGURE 16 1D FIELD EMITTED BY RESOURCES FOR SPECIFIC SERVICES [26] 31

FIGURE 17 PETRI NET MODEL OF THE PRODUCT BEHAVIOR [26] 32

FIGURE 18 INTERDEPENDENCE OF ROUTES [29] 33

FIGURE 19 INTERACTION AMONG GOODS, VERTICES AND VEHICLES IN DLRP [33] 35

FIGURE 20 BASIC LINK-STATE ALGORITHM [34] 38

FIGURE 21 (A) EXAMPLE NETWORK (LINE LENGTHS INDICATED BY NUMBERS BESIDE THE ARROWHEADS). (B)

SHORTEST PATH TREE. (C) ROUTING DIRECTORY [34] 39

FIGURE 22 FORAGING IN BEES [7] 42

FIGURE 23 PSEUDOCODE OF THE CHEMOTAXIS ALGORITHM [36] 44

FIGURE 24 A) EXEMPLARY ITERATION PROCESS, B) CORRESPONDING 𝐴𝑦𝑛 [36] 44

FIGURE 25 TUMBLING PROBABILITY 𝑃𝑘 FOR DIFFERENT 𝑘 VALUES [36] 46

FIGURE 26 CONCEPT OF BMS AT FLOOR LEVEL [38] 47

FIGURE 27 GLOBAL VERSUS LOCAL LEVEL LEARNING [37] 48

(6)

1 Introduction

Autonomous control is all around us, for example house cleaning robots, self-driving cars or welding robots. Autonomous systems are able to make their own decisions. In manufacturing systems, autonomous control is implemented up until a degree where there are almost no employees anymore. Specifically in job shops, autonomous control can greatly improve the performance by solving the logistical problems that arise in job shops.

Job shop manufacturing must cope with ever increasing difficulties. Customers nowadays demand more flexibility in the choice of products and a higher reliability in delivery time and quality. This leads to extremely complex logistics processes. Both the complexity and the dynamic demands make it impossible to come up with the optimal solution for the logistic processes using traditional solutions because there simply isn’t enough computing power to solve them. One way to deal with these difficulties is the implementation of autonomous control in job shop manufacturing. By implementing autonomous control, the logistic problem is split into smaller subproblems. Although this is not likely to give the optimal solution it can give very good feasible results. The benefit of using autonomous control is that such a system can handle the unpredictable dynamics in real world scenarios very well.

There are a lot of autonomous control methods which can be used to solve the logistic problems in job shops, ranging from simple to very complex methods. Some are easier to implement than others and some offer high customizability. But there is a lack of an overview of all these methods, that’s what this report tries to solve.

This report will first discus what autonomous control is and why one would want to use in autonomous control in section 2. In section 3, The job shop will be discussed as well as the job shop scheduling problem. The problem can be solved with different autonomous control methods, these will be discussed in section 4. There are four different categories in which the methods are classified namely: methods that are specifically designed for job shop control, methods that are derived from other practices, methods that are derived from nature and methods that are derived specifically from ant pheromones. All the methods will be explained in this section as well. Some things to keep in mind when implementing autonomous control in job shops will be discussed in section 5. For example, what will happen when different methods are combined. Also, the dynamics in an autonomously controlled system can be of great importance. Finally, section 6 concludes the relevant literature that is found and discussed in this report.

(9)

2 Autonomous Control

To use autonomous control, one must first understand the meaning and characteristics of autonomous control.

The Cambridge dictionary has the following definitions for autonomy and control

Autonomy: “the ability to make your own decisions without being controlled by anyone else” [1]

Control: “to order, limit, or rule something, or someone's actions or behaviour” [2] Together this results in:

Autonomous control is the ability to order, limit or rule something or someone’s behaviour on your own without being controlled by anyone else.

Windt & Hülsmann [3] take the following as the definition for autonomous control:

Autonomous Control describes processes of decentralized decision-making in heterarchical structures. It presumes interacting elements in non-deterministic systems, which possess the capability and possibility to render decisions.

The objective of Autonomous Control is the achievement of increased robustness and positive emergence of the total system due to distributed and flexible coping with dynamics and complexity.

Windt et al. [4] further classify autonomous control in logistics as:

Autonomous control in logistics systems is characterized by the ability of logistic objects to process information, to render and to execute decisions on their own.

Taking together these definition means that for autonomous control in logistics

A set of logistic objects or entities that are structured in a decentralized heterarchical system which can render their own decisions and to take actions to control something.

2.1 Definitions

It is also important to know what specific keywords mean in autonomous control.

2.1.1 Logistic Object

In the context of autonomous control, according to Windt & Hülsmann [3], logistic objects are defined as material items (e.g. part, machine and conveyor) or immaterial items (e.g. production order) of a networked logistic system, which can interact with other logistic objects of the considered system.

(10)

Logistic objects are enabled to detect their situation by processing data from sensors and these objects are also able to assert rendered decisions e.g. to inform a transportation system on a production floor for the transport to another machine.

For example, for a part this goal can be to get work done on itself by a machine. For a machine such a goal could be to handle as many parts as possible or to make the least amount of product changes.

2.1.2 Decentralized decision making

Böse & Windt [5] explain the decentralisation in autonomous control:

One feature of autonomous control is the capability of system elements to render decisions independently. Autonomy in decision-making is enabled by the alignment of the system elements in the form of a heterarchical organisational structure. Therefore, decentralisation of the decision-making process from the total system to the individual system elements is a specific criterion of autonomous control. Each system element represents a decision unit which is equipped with decision-making competence according to the current task. Because decision-making processes are purposeful, according to the decision theory, each system element in an autonomously controlled system is characterised by target-oriented behaviour. Global objectives, for example, provided by the corporate management, can be modified independently by the system elements in compliance with their own prioritisation. For example, the objective low work in process can be replaced in favour of high machine utilization by the machine itself. Thus, the objective system of single elements is dynamic because of ability to modify prioritisation of the objectives over time, i.e. during the production process.

Windt & Hülsmann [3] also discuss the decentralized decision making. They find that decentralization means the delegation of decision power, that is, individual system elements can make independent decisions and can make such decisions by gaining access to necessary resources (e.g. relevant information)

2.1.3 Heterarchical structure

According Windt & Hülsmann [3] heterarchy describes the parataxis of system elements. A Heterarchical system is featured by the absence of a permanently dominant entity. In a heterarchical logistic system such as a production network, there are fewer superordinate and subordinate relationships between logistic elements. This means an increasing level of independence between single elements and a central logistic coordination entity.

2.1.4 Communication between logistical objects

Böse & Windt [5] emphasize the importance of interaction between logistical objects: Decentralized decision-making processes require the availability of relevant information for the system elements. Consequently, the capability of system elements to interact with other is a mandatory condition and thus

(11)

one constitutive characteristic of autonomous control. The ability of interaction can accomplish different values depending on the level of autonomous control. The allocation of data, which other autonomous logistic objects can access, represents a low level of autonomous control. Communication, i.e. bi-directional data exchange between autonomous logistic objects, and coordination, i.e. the ability of autonomous logistic objects to cooperate and coordinate activities of other objects, represents higher level of autonomous control.

According to, Böse & Lampe [6], the autonomy of logistic objects is possible due to recent developments by ICT (information and communication technologies), for example RFID technology (Radio Frequency Identification) for identification, GPS (Global Positioning System) for positioning or UMTS (Universal Mobile Telecommunications System) and WLAN (Wireless Local Area Network) for communication tasks.

Besides the techniques that Böse & Lampe point out, there are a lot of other techniques which can be used for communication such as Bluetooth, Glonas, Galileo, Infrared, Qr-codes and more.

2.2 Reasons to implement autonomous control

Since autonomous control has solved other problems, it could also have a positive effect on logistics. Windt & Hülsmann [3] indicate clearly why autonomous control is considered in logistics:

The drivers supporting the paradigm shift within logistics are categorised in Figure 1 as market, product, technologies and process drivers. The main change, which applies especially to logistic processes, is the significant reduction of time for the change of states, i.e. the time in between two different states of a system. The dynamics within logistic processes are increasing. This may be observed in the categories listed in Figure 1. A heterogeneous market with high demand fluctuations, products which incorporate a high number of variations and have short product lifecycles, new and fast developing information and communication technologies, as well as production on demand, characterise this situation. In parallel, the demands on logistic performance and logistic costs are increasing, too. This is indicated for instance by shorter delivery times, higher schedule reliability delivery flexibility and the use of reconfigurable technologies. As shown in the middle of Figure 1, besides the demands on shorter delivery time, higher schedule reliability, lower price and high quality, the complexity of all the internal and external influencing parameters of logistic systems is also increasing. Among other things, this increased complexity is due to production in global networks, an exponential increase in the amount of data with the use of new ICT, product structures with a high number of variations. In summary, logistic systems are confronted with increasing complexity in combination with many potentially disruptive factors. These impact factors are the drivers of change for a new control paradigm within logistic processes.

(12)

Figure 1 Future conditions and requirements on logistic processes [3]

The paradigm shift is based on the following hypothesis: The implementation of autonomous logistic processes provides a better accomplishment of logistic objectives in comparison to conventionally managed processes despite increasing complexity.

2.3 Characteristics of autonomous control

Two important characteristics in autonomous control are self-organisation (SO) and non-determinism. SO describes a systems ability to organise itself without being controlled by an external entity. Non-determinism describes the effect that in an autonomous control system it is not predictable how the outputs of a system will be regarding its inputs.

2.3.1 Self-organization

Bonabeau et al. [7] discuss the SO of autonomous control systems, they provide the basic ingredients for SO and some key signatures.

The basic ingredients of SO are

(1) Positive feedback (amplification) often constitutes the basis of morphogenesis in the context of this paper: they are simple behavioural ‘rules of thumb' that promote the creation of structures. Examples of positive feedback include recruitment and reinforcement. For instance, recruitment to a food source is a positive feedback that relies on trail laying and trail following in some ant species, or dances in bees.

(13)

(2) Negative feedback counterbalances positive feedback to stabilize the collective pattern: it may take the form of saturation, exhaustion or competition. In the example of foraging, negative feedback stems from the limited number of available foragers, satiation, food source exhaustion, crowding at the food source, or competition between food sources.

(3) SO relies on the amplification of fluctuations (random walks, errors, random task-switching, and so on). Not only do structures emerge despite randomness, but randomness is often crucial, since it enables the discovery of new solutions, and fluctuations can act as seeds from which structures nucleate and grow. (4) All cases of SO rely on multiple interactions. A single individual can generate a self-organized structure such as a stable trail provided that pheromonal lifetime is sufficient, because trail-following events van then interact with trail-laying actions. However, SO generally requires a minimal density of mutually tolerant individuals. Moreover, individuals should be able to make use of results of their own activities as well as of others’ activities (although they may perceive the difference): for instance, trail networks can self-organize and be used collectively if individuals use others’ pheromone. This does not exclude the existence of individual chemical signatures or individual memory, which can efficiently complement or sometimes replace responses to collective marks.

The characteristic signatures of SO include:

(1) The creation of spatiotemporal structures in an initially homogenous medium.

(2) The possible coexistence of several stable states (multistability): because structures emerge by amplification of random deviations, any such deviation can be an amplified, and the system converges to one (among several) possible stable states, depending on initial conditions.

(3) The existence of bifurcations when some parameters are varied: the behaviour of a SO system changes dramatically at bifurcations. For example, pillars built by termites can emerge only if there is a critical density of termites. The system undergoes a bifurcation at this critical number: no pillar emerges below it, but pillars can emerge above it.

2.3.2 Non-determinism

An emergent characteristic of autonomous control is that it is non-deterministic, this is described by Böse & Windt [5]: In accordance with the above-mentioned definition, the main objective of autonomous control is the achievement of increased robustness and positive emergence of the total system due to a distributed and flexible coping with dynamics and complexity. Non-determinism means that despite precise measurement of the system status and knowledge on all influencing variables of the system, no forecast of the system status can be made. Knowledge of all single steps between primary status and following status is not sufficient to describe the transformation completely. Thus, a fundamental criterion of autonomous control is that for the same input and values, there are different possibilities for transition to

(14)

a following status. As already explained, decentralisation of decision-making processes to the system elements leads to a higher flexibility of the total system because of the ability to react immediately to unforeseeable, dynamic influencing variables. In this way, autonomous control can lead to a higher robustness of the overall logistic system. Furthermore, positive emergence is a main objective of autonomous control. Emergence stands for development of new structures or characteristics by concurrence of simple elements in a complex system. Positive emergence means that the concurrence of single elements leads to a better achievement of objectives of the total system than it is explicable by considering the behaviour of every single system element. That means, related to the context of autonomously controlled logistic processes, that autonomous control of individual logistic objects (e.g. machines, parts, orders) enables a better achievement of objectives of the total system than can be explained by individual consideration of the decentralised achievement of objectives (e.g. higher rate of on time delivery, lower delivery times) of each single logistic object.

2.4 Classifying autonomous control

A lot of authors have tried to classify autonomous control or its methods such as Schukraft et al. [8], Windt et al. [9], Böse & Windt [5] and Grundstein et al [10].

Windt et al. [9] classified different control methods and found that for all methods and all logistics target indicators, a relatively robust outcome can be observed. That means that in the presented scenario, the size of the production network does not have a major impact on the performance of the autonomous control methods.

Böse & Windt [5] made a catalogue of criteria to analyse the level of autonomy in an autonomously controlled system. The catalogue of criteria, as seen in Figure 2, represents an easy to use tool that affords an approximate analysis of a logistics system concerning its level of autonomous control. The catalogue of criteria allows both the characterization of an existing as well as a future logistics system concerning its level of autonomous control by determination of the properties of each criterion. Furthermore, two different logistic systems can be compared regarding their level of autonomous control. The last-mentioned point is of importance because this comparison allows an evaluation of the fields of application of autonomy in logistics.

(15)

Figure 2 Catalogue of criteria for analysing the level of autonomy in a logistics system [5]

Grundstein et al [10] designed an evaluation system for autonomous control methods in coupled planning and control systems. The evaluation pattern for complexity enables the definition of production scenarios with different levels of complexity. Using the autonomy evaluation pattern different coupling strategies can be specified and ranked according to their level of autonomy. Thus, every simulation study can be assigned to a unique position in the area spanned by complexity and autonomy. The measurement of the logistic objective achievement will finally allow the building of a provable curve which can be compared to the one depicted in Figure 3.

(16)

Figure 3 Evaluation system for autonomous control methods in coupled planning and control systems [10]

The spectrum of autonomous control methods is way too big to handle all at once so from now on this paper will focus on autonomous control in job shop logistics.

(17)

3 The Job Shop Production System

Job shop production systems are used in cases where the batches are to small to be handled in mass production systems but to large for craft production. In a job shop factory there are a lot of multipurpose machines which can be setup for specific jobs for a specific amount of time. A definition of a job shop is given below by Browne et al. [11]:

Production planning and control in a job shop Batch or job shop production is defined as the manufacture of a product in small batches or lots by a series of operations, each operation being carried out on the whole batch before any subsequent operation is started. The production system must be flexible and uses general purpose equipment to accommodate varying customer requirements and fluctuations in demand. Job shop production is a situation which falls between pure jobbing production and mass production, yet the quantity required is insufficient to justify mass production. Because of the large variety of jobs involved, the job shop operation is inherently complex.

A more general explanation can be found on WhatIsSixSigma.net [12]:

Job shops are designed to manufacture a wide variety of products with small lot sizes to achieve maximum flexibility. Products have usually different operation sequences and operating time for each operation could vary significantly in job shop manufacturing. Products are released to the job shops in batches also called jobs.

General purpose machines are suitable in job shops because they can perform many different types of operations and thus capable of producing a wide variety of products with small lot sizes. Machines which perform similar function grouped together: lathe machines in one department, milling machines in another department and so forth. Job shop manufacturing is shown in Figure 4. A job shop layout is also called a process or functional layout.

(18)

Figure 4 A typical job shop floor [12]

Due to its nature, job shops are very complex. They only produce in small batches and have to be setup specifically for each batch. Mellor [13] explains why job shops are very complex:

The complexities of job shop scheduling are inherent in the nature of the business and are not to be dismissed with a handful of algorithms and a large fast computer. The business is rather more of the nature of a service industry than a manufacturing industry. In general, goods are not produced to be offered for sale, but manufacturing facilities are offered and, almost invariably, products are sold before they are made.

A job shop stays in business if it can satisfy infrequently-repeated demands for a wide variety of highly-specialised products; it can only do this by developing a shop which is an assemblage of " generalised " machines which can be set up in a variety of ways and linked together in a variety of process routes. The resulting complexity of alternative manufacturing processes is a necessary reflection of the variety of products which the shop must be able to make, expeditiously.

These are, perhaps, the circumstances of the pure, or extreme, job shop and all shops are really members of a continuum stretching from such a shop at one extreme to the pure flow of the assembly line at the other. However, the existence of peculiar job shop scheduling problems, in any shop, is largely a function of the prominence of the distinguishing characteristics outlined above. The principal features might conveniently be listed at this point:

 Work is done to meet a customer's order the stock of finished goods is always negative (in inventory control terms).

(19)

 In consequence, there are a very large number of different jobs simultaneously in progress.  Jobs have many manufacturing facilities in common, but their routes through the facilities are

widely different and can be cyclic.

 The times to perform operations on different jobs, or the same job, at the same machine, can vary considerably and the variance of any operation time can be large

Compared to other production systems, job shop manufacturing has some positive and negative sides. WhatIsSixSigma.net [12] lists some advantages and disadvantages of job shops:

Advantages of job shop manufacturing

 Better utilization of machines: In job shop manufacturing, all machines are fully and effectively utilized thus fewer machines are required to manufacture a wide variety of products. Thus, job shop manufacturing needs lower investment because of comparatively less number of machines. Also purchasing cost of general purpose machines becomes lower.

 Greater flexibility: Both a varied quantity and a varied mix of products can be manufactured because there are not dedicated machines. This leads to greater flexibility in job shop manufacturing.

 More effective supervision: As the machines are grouped into departments based on functions performed by them, the specialized knowledge of supervisors about their departments ensures the specialized and more effective supervision. Supervision task becomes more effective because each supervisor must supervise limited same types of machines functioning in his/her department.  Develop high skills & knowledge in operators: Each operator is working in a single department.

This develops high skills and knowledge in operators to perform various tasks in their departments. Management can also take advantages of capabilities of operators to execute manufacturing activities more effectively and efficiently in each department.

Disadvantages of job shop manufacturing

 High material handling cost: Material/parts must be moved from one department to another department. If machine in another department is not ready to accept a job or worker is unavailable, then materials need to be stored and protected there. Status & location of materials are difficult to track and control. Thus, cost of moving, storing, protecting and controlling materials becomes high in job shop manufacturing.

 High material flow distance: When the processing of a part has been completed in one department in the job shop manufacturing, it usually must be moved to next department travelling a large distance. Material flow distance is high in job shop manufacturing because every part may have to travel through the entire shop to complete all the required operations.

(20)

 High production lead time: Each part in a batch must wait for the remaining parts in its batch to complete processing before it is moved to the next stage of production. This causes high production lead time and low production rates in job shop manufacturing.

 High levels of work in progress inventory: Level of work in progress inventory is high in job shop manufacturing because of high production lead time.

 High production cost: A production cost in job shop manufacturing is high because of high material handling cost and high material transportation cost.

3.1 Job shop scheduling problem (JSP)

The problem with job shops is that it is hard to make an efficient and cost effective planning when there are a lot of different jobs. Following the introduction of French [14], Kuhpfahl [15] describes the job shop scheduling problem (JSP): JSP consists of a finite set of jobs 𝐽 = {1, 2, . . . 𝑛} and a finite set of machines 𝑀 = {1, 2, . . . 𝑚}. More precisely, every job consists of a finite set of operations. The processing of an operation must be performed on a preassigned machine, i.e. the 𝑖-th operation of job 𝑗, denoted by 𝑜 , is processed on machine 𝜇 ∈ 𝑀. The operation order of each job is fix, i.e. the technological machine sequence given for every job must be considered. The aim is to find a schedule for processing these 𝑛 jobs on the m machines. Further conditions are as follows:

• The processing of the 𝑖-th operation of job 𝑗 takes 𝑝 > 0 time units. • Each operation must be processed exactly once.

• Pre-emption is not allowed while operations are being processed. • Processing operations may not overlap with one another.

• There is no machine-dependent or sequence-dependent setup time. • Machines must always be available.

Beyond the described structure, the thesis focuses on the standard, dynamic version of JSP with job weights and due dates. In the standard JSP, every job must be processed on every machine exactly once. This means that every job consists of m operations in which every pair of these operations is processed on different machines. This standardization helps to comprise a broad range of problem instances. If a job is not executed on machine 𝑘 ∈ 𝑀 in a problem instance, the corresponding operation is simply neglected in the technological machine sequence. To be consistent with the assumption that every job passes every machine, one can alternatively set the processing time of these virtual operations to zero and add them at the first position of the technological sequence to avoid blockings.

One can furthermore distinguish between static and dynamic JSP. In the dynamic JSP, every job has an assigned release time 𝑟 ≥ 0 so that the first operation cannot start before 𝑟 . In the static JSP, all jobs are

(21)

available at the beginning of the planning horizon, i.e. 𝑟 = 0, ∀𝑗 ∈ 𝐽. Note that the dynamic JSP covers the static JSP.

An additional attribute of a job 𝑗 is its weight 𝑤 which represents the job’s relative importance in comparison to other jobs. Furthermore, every job has a due date 𝑑 ≥ 0 which should, but does not necessarily have to, be fulfilled in a schedule.

The quality of a solution is assessed by the obtained total weighted tardiness, defined as 𝑇𝑊𝑇 = ∑ 𝑤 ⋅ 𝑡 , where 𝑡 = 𝑚𝑎𝑥{0, 𝑐 − 𝑑 } is the resulting tardiness of job 𝑗 in a schedule, and 𝑐 its completion time. The resulting minimization problem is referred to as JSPTWT.

3.2 The real-world problem

Translating the JSP problem to a real-world scenario means that there is a production plant which has different machines which each can handle a lot of different processes or tooling. These machines are used to make custom order parts in small batches. It costs time when a machine must change its tooling. The problem is to plan when which parts will be processed and at what machine.

Scholz-Reiter et al. [16] describes the process that the parts follow in the section below:

At the source the raw materials together with the orders for each product enter the system without any type of WIP control. Each product class has a different production plan, i.e. a list of processing steps that must be fulfilled on the related machines. In case of overload the parts can decide autonomously to change the plan and to use a parallel machine instead. For rendering these decisions, the parts follow predefined algorithms that are called autonomous control methods. The final products leave the system at the drain. See also Figure 5.

(22)

Figure 5 Matrix model of a shop floor [16]

3.3 Solving the problem

There are two ways to solve the JSP, by means of traditional problem solving techniques or by using autonomous control methods. The autonomus control methods are handled in section 4. The traditional techniques are described by Arisha et al. [17] as follows:

Traditional Techniques can be classified under two main categories, i.e. Analytical Techniques and Heuristic Techniques, see Figure 6. The general approach of the analytical methods is to consider the problem in its total system form of scheduling ‘n’ jobs on ‘m’ machines. The relative lack of success of this approach in providing a general optimization method of wide applicability, has led to a switch in the focus of attention from the total system to a more simple decomposed subsystem view of the problem; in which the job shop is considered to be a series of interrelated single machine scheduling problems. Attempts to bridge the gap between heuristic approaches and optimization approach have also been taken. Recently, the Lagrangian relaxation technique has been used to obtain a more efficient enumeration method for a class of JSPs. More recently, a technique to obtain near-optimal solution for parallel identical machines has been used. A shifting bottleneck heuristic as one successful research for decomposing the job shop into sub-problems is presented.

(23)

Figure 6 Traditional techniques [17]

LEKIN Software was designed as a tool with the main purpose of introducing the main scheduling theory and demonstrating the capabilities of several traditional techniques. A concise survey on main Analytical and Heuristic Techniques that have been used to deal with JSP is provided in Figure 7 and Figure 8 respectively.

(24)

(25)

(26)

4 Autonomous control methods for solving the JSP

Figure 9 Overview of the different categories of autonomous control methods “own work”

There are a lot of autonomous control methods, also called algorithms, that can be used to solve the problem described in section 3.2. These methods can be categorized in two different categories, of which one has a subcategory and a sub subcategory. Some methods are specifically designed to solve the scheduling problem in job shops. Other methods are derived from other practices. A lot of these derived methods are inspired by nature and then specifically by ant pheromones. An overview of this categorisation can be seen in Figure 9.

4.1 Methods specifically designed for autonomous control in job shops.

4.1.1 Queue length estimator (QLE)

According to Scholz-Reiter et al. [18], the queue length estimator (QLE) compares the actual buffer states at all the parallel machines that can perform its next production steps. Therefore, the buffer content is not counted in number of parts, but the parts are rated in estimated processing time and the actual buffer levels are calculated as the sum of the estimated processing time on the respective machine. When a part must render the decision about its next processing step it compares the current buffer levels i.e. the estimated waiting time until processing and chooses the buffer with the shortest waiting time. The processing times are not fixed but fluctuate within a frame of 5%. Therefore, the mean values that are used to calculate the estimated waiting time is updated every five days for the QLE.

(27)

4.1.2 Due date method (DD)

The DD method is part of the QLE method, it is compared by Scholz-Reiter et al. [19] and described by Scholz-Reiter et al. [16] as follows:

The DD method is a two-step method. When the parts leave a machine, they use the QLE to choose the machine with the lowest buffer level. Within the buffer the due dates of the parts are compared and the part with the most urgent due date is chosen to be the next product to be processed.

4.1.3 Simple rule based 2 (SRB2)

SRB2 is quite like the pheromone method in section 4.4.1 but it doesn’t include the evaporation part, so all previous data is used instead of only the recent higher quality data. The method is described below by Scholz-Reiter et al. [20]:

Every time a part leaves a machine i.e. a processing step is accomplished, the part leaves information about the duration of processing and waiting time at the respective machine. The following parts use these data about past events to render the decision about the next production step. The parts compare the mean throughput times from parts of the same type and choose the machine with the lowest mean duration of waiting and processing.

4.1.4 One logistics target per rule (OLTPR)

Windt et al. [9] mention the OLTPR method which was developed by de Collaborative Research Center (CRC 637) [21]. The method implements various rules at the machines and parts, where each rule tries to achieve a specific logistics target. It is a method that can be easily extended with new rules to further improve outcome. There are three key rules:

1. Utilization: each machine sends a stronger attraction signal as its buffer becomes less full 2. On time delivery: parts are prioritized by their due date

3. On time delivery: parts prefer machines with short throughput time The various rules must be weighted appropriately to achieve good performance.

4.1.5 Ring like topology (RLT)

Owliya et al. [22] propose a relatively new method below.

The proposed model in this research is based on a ring like topology with an algorithm different to the CNP. This means that resource holons are basically arranged to form a ring as illustrated in Fig. 2. In this model, through the information from order and product holons, a table of tasks to be carried out is created for a manufacturing order. The table includes details and specifications of the tasks, which are prioritized as per the rules specified above. Here, a supervisor holon like the previous model exists. The supervisor circulates

(28)

the prioritized task table in the ring among the resource holons (RHs) successively (like a ring token) and monitors it. Resource holons are sorted in the ring by the rates of their operating cost, which is a known factor for each machine based on its depreciation of investment together with the running costs. The RH with the lowest operating cost receives the token first (for instance, RH-1 as the cheapest resource in Figure 10).

Figure 10 Ring arrangement of the holons [22]

Each resource that holds the token at a given time reviews the tasks remained in the table, and finds the ones that match its technical capabilities. The matching is done in this stage by using if-then inference, which checks for the manufacturing process (turning, milling, assembly, and so on) required by the part and its geometry. Capability of the resource must be higher than, or equal to, what is needed by the part. A resource larger than what is required causes a cost increase. However, the time factor is of highest importance and overrules the increased cost if necessary. The resource then takes out all the tasks that can be performed within their due times with a selfish and greedy behaviour, adds them to its local schedule, and starts performing the one with the highest priority. Each resource cannot be working on more than one task (operation) at a given time. If necessary, selfishness of the resource holons may be moderated toward the overall goal of the system by the supervisor holon. This will depend on the quantity and nature of the tasks set.

Furthermore, the RH leaves a proposal for other tasks that have been unable to be completed prior to their due times. Therefore, the next RH that receives the table, and that is unable to satisfy the due times,

(29)

compares the remaining proposal with that of its own, and decides which to be kept in the task table for further circulation (the worse proposal will be omitted). In this model, the resource holons can interact with all their peers in the ring structure whenever needed (this is shown by the diagonal lines between RHs in Figure 10).

For instance, when a holon has replaced its own proposal for a task, it will notify the holon that had set the previous proposal, to update its local schedule. Each resource has its local schedule, in which tasks’ IDs are saved together with all other attributes of the tasks undertaken, or those for which a proposal is offered. The table will be passed on to the next RH until all tasks are assigned. The logical behaviour of the allocation process mentioned above is shown in the UML sequence diagram of Figure 11.

The solution described above is a new approach to distributed task allocation using a ring structure with advantage of peer to peer interactions. It is completely different to the CNP, although it still uses a bidding mechanism to a limited extent.

(30)

Figure 11 UML sequence diagram for the proposed model [22]

4.2 Methods from other practises

4.2.1 Holonic manufacturing (HOL)

Van Brussel et al. [23] and Márkus et al. [24] describes a holonic manufacturing method. Márkus et al. [24] describe a holonic manufacturing method as follows:

Two types of agents are defined: the first one incorporates a model of managerial activities in the factory, the other one the production related activities. These types of agents are called Management and Machine, respectively. In addition, there is a third party, the Outside World.

The overall objective of the factory is to earn, in a long-range time scale, as much profit as possible. Literally, profit = income - cost. The profit of Management is the payment received from the outside world

(31)

minus the sum of payments given to the machines for working on the orders. The profit of a machine is the difference between payment from the management and its technological cost. Since the whole manufacturing system is considered as a single property, the share of profit among the agents is not considered as a control variable (no mechanism feeds the profit back into production for investments, agents do not go bankrupt, etc.), so the agents' only measure of their acting properly is their profit. To avoid extremities of selfish behaviour and meet the requirements of common-sense manufacturing rationality the agents must obey rules that define a market mechanism.

The outside world submits orders with given time and financial terms. Orders are sequences of tasks. Each task needs a specific volume of a technological resource at the machines. When the Management accepts an order, a job is created. To the tasks the Management attaches time and financial terms, and announces them to the machines. Machines prepare bids for the announcements. Bids accepted by the Management are called assignments see Figure 12.

Figure 12 System configuration [24]

4.2.2 Market based control

Windt et al. [9] describe a market based autonomous control system designed by the CRC 637 [21] and Vollmer [25].

First, a virtual currency is introduced. Parts carry a shopping list of work that needs to be done on them and each job needed for a part has a budget associated with it. The parts auction for access to the machines

(32)

they need for the work on their shopping list. The parts also must consider the distance they travel to the machines as this also has a price.

The algorithm then works as follows:

1. Parts with shopping list and budgets enter the production process. 2. Parts bid on the machines on their shopping list.

3. The highest bidder gets access to the machine.

4. The parts bid according to the minimal price of the machine and the distance cost. 5. Machines grant access for the parts, if they are the highest bidder.

4.2.3 Artificial potential fields (APF)

Figure 13 Interaction between entities in 2D [26]

Vaario & Ueda [27] developed a potential field method for AGV’s. Zbib et al. [26] describe an APF method for job shops: The parts and machines can be translated to respectively products and services. They state that using a potential field approach, allocation and routing problems can be successfully resolved simultaneously. This section has two main subsections. The first introduces the potential field model used in their work. The second describes the behaviour of product used in their FMS.

(33)

Our heterarchical control model using potential fields Modelling the products

Let 𝑃 = {𝑃_𝑖, 𝑖 = 1. . . 𝑁𝑃}be the set of products that move in a manufacturing system to obtain a list of requested services from the set 𝑆 = {𝑆 , 𝑘 = 1. . . 𝑁𝑆}. Let 𝑇𝑌𝑃𝐸 = {𝑇 , 𝑙 = 1. . . 𝑁𝑇}be the set of different available types of products.

Each product Pi is defined by: • its type, 𝑇𝑌𝑃(𝑃 ) ∈ 𝑇𝑌𝑃𝐸;

• its ordered service list, 𝑆𝐿(𝑃 ) = {𝑆 , 𝑘 = 1. . 𝐿𝐿 } (according to its type), where 𝐿𝐿 is the number of services associated to 𝑃 and 𝑆 ∈ 𝑆;

• the service that it is currently trying to obtain from its ordered service list, 𝐶𝑆(𝑃 ) ∈ 𝑆𝐿(𝑃 ). Previous services on the list are assumed to be already obtained, and the following services are assumed to be obtained next.

Modelling the potential fields associated to the resources

Let 𝑅 = {𝑅 , 𝑗 = 1. . 𝑁𝑅} be the set of resources located on the nodes. According to the topology of the system (see the example in Figure 16), each resource 𝑅 emits attractive potential fields in 1D for the different services available on the resource. The value of the field decreases with the distance and propagates in the opposite direction from the product flow.

Products make allocation and routing decisions at specific locations, called decisional nodes 𝐷 , corresponding to divergences on the routing map. The resources’ attractive fields are only read (i.e., sensed) on these decisional nodes.

𝐹 𝐷 , 𝑅 , 𝑆 denotes the attractive field value diffused by a resource 𝑅 for the service 𝑆 to attract a product located at decisional node 𝐷 at time 𝑡. The field formulation is independent of the product type because the resources do not know to which product they emit their fields. The attractive field value sensed on a decisional node 𝐷 depends the attractiveness of the resource 𝑅 offering the service 𝑆 at time 𝑡 and must be modified according to the distance (spatial/temporal) between 𝐷 and the location of the resource 𝑅 .

 Attractiveness of the resource at time 𝑡 𝐴𝑇 𝑅 , 𝑆 This first parameter considers the quality of a service 𝑆 on the resource 𝑅 , the resource’s workload and its availability (e.g., whether it is undergoing maintenance). This parameter, which is set at the location of the resource 𝑅 , corresponds to the maximum value of the attractive potential field. This maximum value will evolve over time (i.e., increase or decrease) with the resource’s workload.

(34)

o Quality of the services on the resource 𝑄 𝑅 , 𝑆 denotes the quality of the service 𝑆 on the resource 𝑅 at time 𝑡. The greater this value, the more attractive the resource. o Availability of the resource 𝐴 𝑅 is a Boolean value set to 1 if the resource is available at

time 𝑡; otherwise, it is set to 0 (e.g., in cases of breakdowns or maintenance operations). To avoid deadlocks, this value can be set to 0 if the maximum capacity of the waiting area has been reached.

o Current workload of resource Let 𝐿𝑊𝑄(𝑅 ) be the maximal length of the waiting queue; 𝑁𝑊𝑄 𝑅 , the number of products in the queue at time 𝑡; 𝑊𝑄 (𝑅 ), the 𝑖 product in the waiting queue (FIFO); and 𝑊𝑄 (𝑅 ), the product currently being processed. The current workload of a resource depends on two parameters:

 𝑇 𝑅 , 𝑊𝑄 𝑅 , the time remaining until the product currently 𝑊𝑄 (𝑅 ) being processed on 𝑅 at time 𝑡 obtains the requested service; and

 the workload associated to the products in the waiting queue. The greater the workload of the candidate resource, the less attractive the resource.

 Decrease in the attractive field with distance. The attractiveness field changes with the distance between the product location Di and the location of the resource R j emitting the field. Several solutions for decreasing this field exist; these solutions mainly use decreasing exponential, linear or inverse functions. In this paper, a linear attractiveness of 𝑅 over time function was used as a first attempt; with the aim of evaluating the influence of the possible types of decrease a posteriori (after real experimentation) to accurately characterize the impact on the global performance.

(35)

Figure 14 Evolution of the attractiveness of 𝑅 over time [26]

In Figure 14, the double lines show the variation over time of the attractiveness for a resource 𝑅 offering a service 𝑆 . At time 𝑡 , 𝐴𝑇 𝑅 , 𝑆 = 𝑄 𝑅 , 𝑆 , with 𝑅 being available and 𝑁𝑊𝑄 𝑅 = 0. At time 𝑡 > 𝑡 , the workload of 𝑅 is greater (e.g., products are waiting in the queue), making the resource less attractive. At 𝑡 > 𝑡 , the attractiveness value is set to 0 because 𝑅 is unavailable (e.g., due to a maintenance operation). At 𝑡 , 𝑅 again becomes available, but its attractiveness value depends on its workload. The grey curve in Figure 14 shows the decreasing value of the attractive field diffused by 𝑅 for the service 𝑆 at time 𝑡 with distance 𝑑.

This distance is fundamentally spatial but will be associated to a temporal value for two reasons. First, a finite speed for each product is considered, which will affect the travel times. Second, travel times will evolve according to the workload of the transportation system and its availability (e.g., possible traffic jams in the transportation system). Nonetheless, the physical distance will be used to set the initial values of the decrease function. To do this, as the FMS can be described as a strongly connected graph, the distance between two points is not the Euclidean distance but is the sum of the lengths of the arcs connecting these two points. Thus, in the example shown in Figure 15, the distance between the starting node 𝐴 and destination node 𝐸, can be expressed as 𝑑(𝐴, 𝐸) = 𝑑 + 𝑑 + 𝑑 + 𝑑 , where 𝑑 is the length of each arc 𝑖 traveled.

(36)

Figure 15 An example of a distance calculation [26]

The distance shown in Figure 15 can be calculated by an optimization algorithm (e.g., the Dijkstra algorithm) since there is circuits, and the arc values are all positive. Let 𝐷 = {𝐷 , 𝐷 , . . . , 𝐷 , . . . , 𝐷 } be the set of decisional nodes. The distance matrix at the time 𝑡 is:

𝑀 𝑑 = 𝑅 … 𝑅 … 𝑅 𝐷 … 𝐷 … 𝐷 _⎝ ⎜ ⎛ 𝑑 … 𝑑 … 𝑑 … … … … … 𝑑 … 𝑑 … 𝑑 … … … … … 𝑑 … 𝑑 … 𝑑 _⎠ ⎟ ⎞

where (𝑑 )t is the optimal transportation time between the decisional node 𝐷 and the resource 𝑅 at time 𝑡. This matrix will be re-estimated using real measured transportation times. Updating this matrix will cause the attractive fields at each decisional node Di to automatically adapt.

Choosing the resource

The current product 𝑃 seeks to obtain the current service 𝐶𝑆(𝑃 )(corresponding to service 𝑆 ) in its service list. When arriving at a decisional node 𝐷 , the product must make an allocation/routing decision to find the most appropriate resource. The decision is based on the intensity of the different attractive fields sensed at location 𝐷 . Since each resource emits several fields simultaneously for each possible service, products may apply a filter to select only fields concerning the service 𝑆 they want to obtain. Since according to one of their assumptions, the operating time for one service and for one kind of product is same for all resources, it is not necessary to take this time into account in the decision.

The resource is chosen at time 𝑡 using the following relation, meaning that the maximum value is sought:𝐹∗_{(𝑆 ) = max 𝐹 𝐷 , 𝑅 , 𝑆} _.

This choice is made in a deterministic way according to the following rule: 𝑅∗_{= arg 𝑚𝑎𝑥} _{𝐹 𝐷 , 𝑅 , 𝑆} _.

(37)

Figure 16 translates the principle of the 2D field in Figure 13 by applying it to a 1D graph representing the production system. This figure shows three resources; each resource emits a potential field for a specific service, respecting the topology of the graph conveyor: R1 (service S1), R2 (service S2), R3 (services S1 and S3). A product enters in the system with a service list (S1, S2 and S3) and, respecting the graph conveyor, moves to the correct resource to obtain its services in sequence. In the example in Figure 16, the product selects S1 as its first service; it is attracted by R1, which can render this service. Since the services are ranked, if the sequence of services changes, the product behaviour will not necessarily be the same. In other words, the product may not choose the same resource nor chose the same path to reach the chosen resource. However, the same product will be obtained regardless of the product’s choices, with perhaps a different length of time.

The choice of a destination by a product can be seen as one action in one state. Other states must be considered, such as moving or waiting for service. These different states and their relationships are explained in the following sub-section.

Product states

The Petri net in Figure 17 provides the state/transition graph model for the products. Interpreting this graph highlights the different steps:

• Step 1—After being put in the system, the product must obtain all ranked services from its service list. At the beginning, the current service is the first service on the list.

• Step 2—The product searches for a service node able to satisfy the current service on its list. To reach this node, the product moves to a decisional node to decide which routing arc it will take. To do this, it reads updated values of the potential fields filtered according to the type of service requested.

• Step 3—If a product senses no field, meaning that none of the resources can provide the requested service, the product will choose the next node that will enable it to reach the reference loop. It will then remain in this loop until it senses an appropriate field (i.e., until at least one resource is available for the requested service), which allows deadlocks to be avoided. If a product senses a field, it chooses the service node corresponding to the highest value (as explained previously). • Step 4—The product moves toward the chosen service node. It will reach the service node where

resource is located.

• Step 5—Once the product is at the service node, if this resource is busy, the product enters in the waiting queue. If the resource is free, the product is processed (i.e., it obtains the desired service). • Step 6—After processing, the product sets the current service to the next service on its list, and

(38)

• Step 7—When the services list has been completed, the finished product moves toward the system output.

Figure 17 describes the nominal behaviour of product.

Pach et al. [28] later improved the APF method to overcome some limitations.

(39)

(40)

4.2.4 Distributed logistics routing protocol (DLRP)

The DLRP is based on how the internet works and is an implementation of physical internet. Wenning et al. [29], Wenning et al. [30], Rekersbrink et al. [31], Rekersbrink [32]and Scholtz-Reiter [33] all discuss the DLRP. Wenning et al. [29] give the following description of the DLRP.

In the following, a concept for distributed routing in a logistic network is presented. In this concept, vehicles as well as packages are considered as autonomous. They have sufficient intelligence and communication capabilities to get their information and to decide on the next steps to be undertaken.

Figure 18 Interdependence of routes [29]

In this concept, next steps mean calculating a route or deciding about being loaded into a vehicle (from the package’s view) or picking up a package (from the vehicle’s view). If both the vehicles and the packages determine routes based on their individual goals, the dilemma arises that the routes are most probably different. To make it worse, the decisions are interdependent: The package needs knowledge about vehicle routes to find candidate vehicles and the vehicle needs knowledge about the package routes to be able to find an efficient route where its capacity is best utilized. Figure 18 illustrates this interdependence. The interdependence implicitly gives rise to another issue: The knowledge of each other’s existence, i.e. how does the package know which vehicles are there, and further: How does the vehicle know about the packages? If there is no way to get to know about each other, they cannot communicate and thus cannot exchange their information.

(41)

Direct communication: An entity, say a package that enters the system, broadcasts some information about itself and collects responses from all other present entities. This is very inefficient and would lead to a high load of communication signalling, and the entities which are currently out of communication range might not get the information.

Indirect communication: This assumes the presence of knowledge brokers or repositories in the network. In this way, both the vehicles and the packages know entities to whom they can send their information and where they retrieve other information.

Due to the drawbacks of the other solution, the indirect communication was chosen as the way to solve the interdependence problem. As it is not intended to introduce an additional central repository, which would in fact foil the idea of a distributed system, the vertices that are present in the logistic network are chosen as the “relays” for indirect communication and therefore as the knowledge brokers. This fits perfectly into the distributed nature of the concept, as each vertex has only a part of the global knowledge, rather than the complete knowledge about all routes and all packages in the system.

The vertex is a knowledge broker for the vehicles and packages. Before deciding about a route, a vehicle/package requests current information from the current or next vertex. Each vertex includes relevant information available from its current knowledge-base and forwards the request to neighbour vertices. The neighbour vertices do the same and forward it further. This way, the request is propagated through the network until the destination or a predefined hop limit is reached. Then the last vertex creates a reply message that is sent back directly to the originator of the request. This reply contains all the information that has been collected during the propagation of the request message through the network, including the last vertex. In general, an entity can receive more than one route reply as there are multiple paths possible. As it is not known how many replies would get back, a timeout and an upper limit for the number of replies are specified to trigger the decision process without long waiting periods.

After receiving the reply messages, the entity is ready to make its route decision based on its individual preferences and the data received. After making the decision, it withdraws its old route if any, and announces its new route to all relevant vertices. This way, the vertices get an information update, which will be used in processing the future requests. Figure 19 shows the information flow in DLRP.

(42)

Figure 19 Interaction among goods, vertices and vehicles in DLRP [33]

This approach also leads to uncertain knowledge: As a package does not know in advance whether a specific vehicle picks it up or not, it looks for a set of alternative routes to increase the probability to reach its destination in time. All these alternative routes are announced to the vertices, so that the announced package routes are just valid with a certain probability. If a package is picked up by a vehicle, unused routes must be cancelled again. Vehicles on the other hand do not necessarily stick to a single route, so the vehicle routes also are uncertain. The vehicles check the current state of their options whenever they reach a vertex. If they find a route that is better than the original one, they can either change their decision depending on their individual settings, or stick to the old one.

The DLRP itself does not specify the functions that are used by the packages and vehicles to decide about their routes, it just specifies the interaction. Therefore, it should be regarded as an interaction framework which provides a basis for distributed information management and decision making in logistic scenarios. The logistic performance that can be achieved with this framework strongly depends on how the logistic entities utilize the information they can obtain. There are several possibilities for decision making, for example fixed rule sets (e.g. always take the shortest route), heuristic, probabilistic or fuzzy logic approaches etc. Some of these options are under investigation for their use in the DLRP framework.

4.2.5 Link-state internet routing protocol (LSIRP)

Windt et al. [9] quickly explain the LSIRP, developed by CRC 637 [21] as follows: 1. A map of the facility and the connections between machines is built/provided

2. Shortest paths are computed based on various chosen criteria, generally using Dijkstra’s algorithm 3. As the situation changes (breakdowns, buffer states, new machines) only the changes are

propagated among the machines

Autonomous control methods in job shop logistics - Methoden voor autonome controle in job shop logistiek

Specialization: Transport Engineering and Logistics

Report number: 2017.TEL.8182

Title:

Autonomous control methods in job

shop logistics

Author:

R.O. de Boer

Summary

Table of figures

Contents

SUMMARY

I

TABLE OF FIGURES

III

1 INTRODUCTION

1

2 AUTONOMOUS CONTROL

2

2.1 D

2

2.1.1 L

O

2

2.1.2 D

3

2.1.3 H

3

2.1.4 C

3

2.2 R

4

2.3 C

5

2.3.1 S

-

5

2.3.2 N

-

6

2.4 C

7

3 THE JOB SHOP PRODUCTION SYSTEM

10

3.1 J

(JSP)

13

3.2 T

-

14

3.3 S

15

4 AUTONOMOUS CONTROL METHODS FOR SOLVING THE JSP

19

4.1 M

.

19

4.1.1 Q

(QLE)

19

4.1.2 D

(DD)

20

4.1.4 O

(OLTPR)

20

4.1.5 R

(RLT)

20

4.2 M

23

4.2.1 H

(HOL)

23

4.2.2 M

24

4.2.3 A

(APF)

25

4.2.4 D