Interaction in hybrid intelligent scheduling - CiteSeerX

International Journal of Human Factors In Manufacturing, 6(3), 185-203.

Interaction in Hybrid Intelligent Scheduling Peter G. Higgins School of Mechanical and Manufacturing Engineering Swinburne University of Technology, Hawthorn 3122, Australia ABSTRACT Hybrid human-computer intelligent systems for scheduling have evolved in response to the inability of algorithmic methods to handle the complexity of production in real manufacturing environments. They combine the abilities of humans to recognize patterns in data and to make inferences with computer methods for decision making. The paper discusses (1) the factors influencing scheduling decisions in small-batch manufacture and the role of humans in the scheduling process, (2) the position of the human scheduler in hybrid intelligent decision-making processes, and (3) the inadequacy of using Gantt charts as the standard interface for human-computer interaction in decision making. The paper proposes, for humans to be active partners in decision making, the primary interfaces should display detailed characteristics of jobs in a way that reveals patterns in the data and helps inferential processing.

1.

INTRODUCTION

Discussions on interactive scheduling have paid scant attention to the special characteristics and particular needs of job shops in environments that are nowhere near the leading edge of Automated Manufacturing Systems. Many small shops have the following characteristics: • small to medium lot sizes • operation times that are short • customers who place orders with short lead time. Such shops have little scope to make long-term plans. Systemic complexities also force scheduling activity to be directed to very short time-horizons. Because resources are limited, they usually find it too expensive and difficult to develop a comprehensive data base that can furnish all information used in making scheduling decisions. In this article, the distinctive characteristics of small job shops that make schedule construction difficult are discussed. Drawing upon a case study, features are proposed for “hybrid” human-computer scheduling systems that meet the needs of small job shops.

2.

REAL SCHEDULING ENVIRONMENTS

In recent years, researchers have sought ways to overcome the shortcomings of the classical operations research approach to job-shop scheduling. The algorithms and scheduling systems developed for the classical n/m problem are of little practical value (Jackson and Browne, 1989). Simplifying assumptions, intended to remove the computational complexity, also make the problem less relevant to actual practice. Shops are perceived as deterministic. Manufacturing processes are assumed to be simple, stable and well understood with predictable and reliable set-up and processing times, known delivery quantities, times, and qualities. This characterization is far from the shop-floor reality for most cases (McKay,

1

Higgins, P.G. (1996) Interaction in Hybrid Intelligent Scheduling. International Journal of Human Factors In Manufacturing, 6(3), 185-203

Safayeni and Buzacott, 1988). In scheduling, theoretical methods are the antitheses of the praxis: their goals differ; they use different information; and they observe different practices. Schedulers cannot ignore the predicaments of the real world. Work environments are generally unpredictable. Unplanned events occur frequently during the scheduled period. People often have to schedule in circumstances where procedures are ad hoc, records are kept on paper, and decisions are made that are almost arbitrary and outside the control of the scheduler (Solberg, 1989). Schedulers often do not adhere to classical assumptions in the strategies they pursue. They: • assign priorities to jobs • change the size of batches • split operations between machines and overlap operations to speed up work • interrupt operations to run more urgent jobs • renegotiate due dates with customers to spread the work load, and • use machines in non-standard ways to increase short-term capacity. New jobs usually arrive before previously scheduled jobs in the system have been processed. The new arrivals may make the prevailing plan irrelevant. The state of the shop may restrict choices available for amending the schedule. For example it may be impractical to alter the place of some jobs in a queue. Often those jobs for which processing is imminent have already placed calls on resources and materials. Reversing these calls may be difficult. Also, frequent changes to the processing order of jobs that are currently on the shop floor may cause chaos and confusion. To limit these effects, a scheduler may decide to restrict changes. Multiple goals, frequently in conflict, beset schedulers. The conventional OR (Operations Research) approach is to construct a single aggregated function for measuring performance. Each goal is assigned a weight based on its relative importance. Then, a linear function, formed by summing the weighted goals, is pursued. However, schedulers often apply goals that may be neither clear nor explicit. They can follow goals without having to articulate them. They pursue practices they believe to be good based on years of experience with the production process. In his field study McKay (1987) found that only a few per cent of constraints and issues used by schedulers in decision making are normally supplied facts. The remainder were semantic relationships requiring inference and induction. A key factor in scheduling is uncertainty. It dominates scheduling activity. Uncertainty extends beyond the usually accepted sources (i.e., arrival times, scrap rates and machine breakdowns). McKay, Buzacott and Safayeni (1989) found that these other issues, which are myriad, are more important. Furthermore, “Almost any factor, constraint, or goal can have variety and certain combinations of variety may or may not have effects on other constraints or goals.” Constraints may include environmental and seasonal conditions, transportation, raw material, type of work, labor force, and labor rates. They stress that not all constraints and goals are active simultaneously. Goals depend upon the hour or the day, and constraints change: “what is a ‘good’ schedule generated Monday morning may be considered to be a ‘bad’ schedule if generated Monday afternoon.” These practices often depend upon schedulers having deep knowledge of the working environment. In seeking to advance scheduling practice this tacit knowledge should not be disregarded. One way to harness this knowledge is through AI (artificial intelligence). First, a knowledge engineer needs to collect all the rules a scheduler has in his/her repertoire. Then, a 2


constraint-based reasoning system applies these rules to produce feasible schedules (Fox and Smith, 1984). For the system to be effective — that is, to operate without human intervention — it has to cater for all factors and circumstances that could ever arise. Alas, this is its Achilles heel. To be fully automated, the system has to cater for all circumstances that could ever arise. It may have to cater for the presence of multiple resources and routings, operation precedence, job priorities, random failure, availability of material, changes in production goals, and calls to expedite some jobs. While rules can be written for many of these factors, their capture is usually costly. Companies, especially those that are small, are reluctant to expend sufficient time and money on creating rule bases that are comprehensive enough to make well-formed decisions. Problems arise even for systems that initially perform well. What happens to the rule base when resources change, products change, and new methods and materials are introduced? Does someone think about upgrading it, or, is nothing done and the subsequent errant advice acted upon without question? Has the company retained employees who have a detailed understanding of scheduling? Has the company access to persons with intimate knowledge of the rule base? This is particularly problematic. Maintaining this knowledge within the organization is expensive. Alternatively, by purchasing outside expertise the company becomes reliant on another enterprise. This agency may not maintain its knowledge of the scheduling system as personnel, business interests and activities change. An alternative approach is for humans and computers to make decisions interactively. In recent years interactive scheduling systems have been presented as a way forward. Human and computer intelligence combine to solve the scheduling problem (Nakamura and Salvendy, 1994). The human handles the unstructured part of a scheduling problem. For the more structured parts of the problem the computer supports the human. For example, the computer can apply procedures for scheduling jobs from a set of standard heuristics — First Come First Served (FCFS), jobs with shortest processing time first (SPT), etc. In the interactive process the user exercises local knowledge. This knowledge is not restricted to quantitative information (e.g. order size, due dates). It also includes subjective information about customers, operators and management priorities. Furthermore, humans can cope with information that is incomplete, ambiguous, biased, outdated, and erroneous (McKay et al., 1988).

3.

INTERACTIVE SCHEDULING

The form interactive scheduling may take is wide ranging. Sheridan (1980) offers ten variants in the sharing of responsibility between humans and computers. At one extreme, the human acts as the principal controller, taking advice from the computer. The opposite bound has the computer as the principal controller with the human performing corrections and adjustments (Sanderson, 1989). Nakamura and Salvendy (1994) locate hybrid intelligent systems between these extremes. In their definition of hybrid intelligent systems, both computer and human put forward solutions. Computers also carry out data processing, displaying results graphically. Importantly, proposed schedules are evaluated and modified by humans. They see the primary role for humans as achieving total-system performance objectives. The components of Nakamura and Salvendy’s decision-aiding system for an FMS (Flexible Manufacturing System) are a human performance model, an interface, a human scheduler, a schedule generator, and an FMS model. The human performance model captures scheduling

3


behavior of the human scheduler using a variety of knowledge sources that include heuristic algorithms, optimizing procedures, and rule-based procedures. Another formulation of a hybrid system is the integrated approach of McKay et al. (1989). Their system consists of a knowledge base, an expert system, a schedule generator, a domain manager, a schedule modification system and a human scheduler. As the human scheduler enhances and tunes the schedule, the schedule modification system “learns”. The human scheduler refines the schedule and provides feedback about the current situation to the schedule modification system. Both models include the means for the computer system to capture human scheduling behavior: for McKay et al. it is the schedule modification system and for Nakamura and Salvendy it is the human performance model. Both these research groups present the gamut of abilities that humans bring to the scheduling task. Of these, the identification of patterns and clues, and the provision of supplementary direction for constraint relaxation, are important to the following discussion. Interactive scheduling systems generally use Gantt charts as their primary interface. The computer first builds a schedule and then displays it as a Gantt chart. Depending upon the system, the user may select from several choices or the user may modify a chart by moving jobs on the screen (Bauer, et al., 1994). Moves violating hard scheduling constraints (e.g. operation precedence) are normally barred. Frequently, a rebuild-algorithm then reschedules activities that temporally follow these manual changes. The interplay between human and computer continues until some satisfactory schedule evolves. In this formulation of interactive scheduling humans are subsidiary. This may lead to decision making being biased inappropriately towards the computer’s determination. Interactive systems that apply OR heuristics normally consider only a few, at most, of the available job attributes that may have some bearing on the schedule’s performance (Nakamura and Salvendy 1987; Hwang and Salvendy 1983, 1988; Tabe and Salvendy 1988; Tabe, Yamamuro and Salvendy 1988). Hybrid intelligent scheduling systems make use of knowledge bases to redress these deficiencies. As the system learns from the activities of the human scheduler, some problems in establishing a rule base, described above, are sidestepped. Because computers and humans put forward solutions in partnership, unlike automated systems, it is not critical for the knowledge base to be fully comprehensive. For humans to be actively involved in decision making, the computer system needs to support their decision-making processes. Humans reflect upon the characteristics, that is, the attributes, of jobs and the calls that these make upon the shop. The job attributes, and patterns among attributes across jobs, act as stimuli. They relate the characteristics of the job to the state of the working environment. Using their deep knowledge of the domain, they draw inferences about possible scheduling strategies. In this reflection they may consider, for example, the following issues (McKay, et al., 1989): What goals currently have high priorities for management? Which jobs need immediate attention? Which jobs may be left to later? Which jobs are likely to cause problems? What processes are difficult to repeat or set up? What material or quality issues will arise? How can the special competencies of humans in pattern recognition and inference be supported? Humans need to be at the core of the decision-making process, instead of being only involved in the alteration of a schedule a computer has built. As the sources of uncertainty are many, and scheduling factors in real manufacture are myriad, much of the decision making has to be left to humans. A Gantt chart is a display of output. It is the product (a plan for running the shop) of a decision-making process. Gantt charts only show partial information about the jobs: 4


commonly, job numbers, customer names, processing times and sometimes due dates. Humans taking responsibility for the production schedule need to make well-informed decisions. Displaying results is quite a different activity to using a display to search for inferences. The difference is communication as against discovery. A Gantt chart’s purpose is to communicate the plan of work to people who have to use it. The imperative is simplicity (Bertin, 1981). Whereas, a display for decision making must comprehensively show all data a scheduler may use to discover relationships. It has to reveal all the relationships formed by the interplay of the data. Therefore, a Gantt chart should not be the principal means for humans and computers to interact. Some advanced systems provide graphic representation of the plant as an additional aid for the scheduler. A user can scan the state of the shop in real or simulated time. While showing the state of the shop is advantageous, schedulers still need a display in which they can observe the characteristics of the jobs that may affect their scheduling decisions. Human schedulers tend to approach their task in an opportunistic way (Woods and Roth, 1988). This requires an interactive system in which the interface elements, and their location, are not tightly bound to a restrictive perspective of the problem. The architecture of a hybrid intelligent system developed by the author is shown in Figure 1. Two types of display — the Jobs Screens and the Gantt Chart — allow the human scheduler to interact with the computing system. There is a separate Jobs Screen for each machine and one for jobs yet to be allocated to a machine. During decision making, the human scheduler focuses primarily on the Jobs Screens. Meanwhile, the scheduler keeps an eye on the current state of the Gantt chart to see when jobs are expected to be loaded and completed. A scheduler can see the values of the job attributes in the Jobs Screens. If all the job attributes that may affect a schedule can be display, the scheduler could regard all dependencies and conceivable interactions. While many attributes may be in the plant’s database, schedulers may also use other factors — especially those that are unstructured. If the system allows them to annotate the record with comments, it could remind them of these other significant characteristics of the job: the importance of the customer and expected difficulties with the job. In allocating jobs to machines, and arranging the processing sequence, the human interacts with the computer’s knowledge-based adviser. The scheduler may select jobs and arrange them according to one or more rules from the computer-supplied OR heuristics. By grouping and observing jobs, the human infers what factors are important for the current set of jobs. Unstructured factors such as experience, knowledge and intelligence of the person using the system may be important (Bauer, et al., 1994). The scheduler allocates jobs to machines and sets the order of processing by moving them about the screen. Eventually an acceptable Gantt chart issues from this interplay between the computer and human. Having the human central to the decision-making process allows development of a truly generic system. In its most basic form, the human makes all the decisions. The computer provides tools for ordering jobs to standard OR scheduling rules. Either through the system learning the rules the human applies, or, by production rules being directly entered, the knowledge-based adviser, or equivalent, may in time be able to offer meaningful advice.

4.

CASE STUDY

A tangible example may help make the issues, raised about practical aspects of job-shop scheduling, more lucid. The behavior of a scheduler at a printing company, Melameds for the

5


sake of discussion, has been intensively studied. Melameds prints forms on continuous fan-fold paper for such uses as checks and invoices. The scheduling environment is typical of many small shops: • Some-of-the-machines are the state of the art, but they are not in the realm of Advanced Manufacturing Technology • There is no on-line monitoring of processes • The details of jobs are maintained on a plant-wide computerised database. Currently, the scheduler only has a machine-loading board to aid him. It has the names of the machines written across its top. Cards, listing the details of jobs, slide in vertically laid tracks. He places each card under the machine needed for its next operation. Cards are arranged vertically in the order that they are to be processed. At the beginning of a shift the scheduler tries to project a schedule across two shifts. As the shift progresses, the scheduler regularly modifies the plan in response to unplanned events. The machine-loading board is an inadequate aid. In observing the board, the scheduler cannot see: when a machine will be idle; how long a job requires for processing; nor when it will leave the shop. The board is not just an ordered list of job operations waiting to go onto each machine. It is therefore unlike the boards Gibson and Laios (1978) used in their comparison of board displays for job-shop scheduling. Each card displays the characteristics of the job that could influence scheduling decisions. The cards are color coded to highlight the size of the printing cylinder: the item that takes the longest time to change on a press. To keep the case-study simple, we shall only consider the presses, the primary machines, and will ignore the downstream processes. The problem thereby reduces to scheduling jobs that have a single operation. The presses are the same, except the number of colors (one, two, four and six) they can print and the ancillary attachments, providing extra functionality, differ. The ancillaries place additional constraints on the allocation of jobs to presses. For instance, all jobs using ink that needs ultraviolet fixing must go to the six-color press. This manufacturing environment, while not being extraordinary, is a member of the difficult class of an identical parallel-machine problem (French, 1982). The number of parallel machines a job “sees” depends upon the number of colors used in its production (see Table 1). Over recent years there have been some theoretical advances in allocating and sequencing jobs on parallel machines with major and minor set-ups (Tang, 1990; So, 1990; Rajgopal and Bidanda, 1991; Wittrock, 1990). While these advances move forward the understanding of this class of problem, they apply to situations that are far simpler than this case-study (in which the level of difficulty is nothing unusual). To satisfy his primary strategy to maximize machine utilization, the scheduler at Melameds tries to keep the four presses operating productively. The principal way to achieve this is to minimize time wasted in setting up machines (Higgins, 1992). At Melameds three job attributes, depth, color, and width, affect the set-up time: Depth: The size of the printing cylinders must be an exact multiple of the depth of the form. Cylinders take 40 min to change. Width: The perforations between sheets cannot extend to the paper’s edge. Where the width reduces between jobs, the operator can quickly break teeth off the perforating

6


tool to stop holes being punched. For increasing width, the operator must first replace the perforating tool and then remove teeth near the edge. This takes about 7 min. Color: Washing out and replacing a color takes 10 min. The processing time for most jobs is between 20 min and 4 h. For small jobs, all these factors contribute significantly to lost time. At the upper end of the range, wash-ups and changes to the perforating tool only make a minor contribution to the total time for an operation. Hence, in scheduling large jobs, the scheduler may willingly accept otherwise avoidable changeovers so that he may meet other objectives. The scheduler’s strategy is more complex than ordering jobs merely to minimize set-up time. In choosing a schedule, he deliberates upon other aspects of the job and the environment. His goal is composite. It includes, but not solely, the maximization of machine utilization and the minimization of tardiness. Of course, these can be expressed as a weighted linear function. However, this is not the scheduler’s practice. At Melameds, among other objectives, the scheduler’s desire is to maximize machine utilization and minimize tardiness. If the problem somehow was reducible to these quantifiable goals, a single indicator made up of their weighted sum could be formed. Each goal places a different demand on the shop. These demands conflict. For example, minimization of tardiness requires ready availability of machines. This implies machine utilization is not maximum. Therefore, the performance indicator is highly sensitive to the value of the weights. Being a real shop the relative importance of each goal is not immutable. It may vary with the time of the year, month, or week, the shift, the value of the current jobs, customer goodwill, and the practice of competitors. The more sophisticated hybrid intelligent systems attempt to capture the complex behavior of humans in scheduling in such circumstances. The computer system then can anticipate what the human would consider as an appropriate action. The scheduler also has to consider other, less tangible, elements of the goal, for example satisfying customers. Satisfaction is difficult to express. It does not have a singular meaning. For some customers the meeting of the date agreed upon is most important. For others, the primary interest is for the turnaround to be fast, while other customers are concerned about the quality of the job. At Melameds the highest quality is only achievable on the six-color machine running at less than normal speed. These restrictions contravene the conditions for maximizing machine utilization and reducing average turnaround time. Colors affect scheduling in two ways. If colors change between jobs, washing the color applicators may be necessary. A scheduler’s choice of what to do is complex. As an example, consider a three-color job printed on a six-color press. The job that follows has completely different colors. Time will not be lost to cleaning if the three unused applicators are clean or have in them inks of the required colors. This action delays cleaning. The four-color machine can also produce the three-color job, leaving the six-color machine available for jobs requiring five or six colors. The scheduler’s choice will depend, inter alia, on the current sequences on both machines, the possible sequences on each, the set up costs on each when the job is loaded and the total set up cost for each over a selected period. As the company’s strategic niche in the market is the provision of quick turnaround, the scheduler tries to have cylinders for the most common depths immediately available. He can then process a premium job without waiting 40 min to change over cylinders. This rule is not simple. The obvious configuration is for the cylinder size with the greatest demand from such jobs being on the six-color press to allow processing of all jobs of that size. This may also decrease the time to wash the applicators. When the likelihood of “quick-turnaround” jobs for

7


the other cylinder sizes is considered, along with some estimate of the number of colors to be printed, the allocation of cylinders to machines becomes more problematic. The obligation to process existing jobs, efficiently and on time, tempers the scheduler’s desire to provide quick turnaround for jobs that may never exist. Correctly reading the meaning for some factors depends upon context. For some customers the due date is rigid. Others may not be too concerned if they receive their orders a day or two late. Yet again, these very same customers may have jobs in the system with due dates that are atypically firm. The attitude towards the customer also influences the significance placed on a due date. The disposition of the scheduler, manager, or sales representative towards a customer who is regular, new, slow to pay, belligerent when jobs are late, etc., may affect what delivery date Melameds consider acceptable. Different departments or persons may see particular constraints quite differently. A sales representative, a production supervisor and a customer may hold quite different views on the firmness of a due date. While a customer may not be unduly concerned about a late delivery, a sales representative may see a late delivery as a threat to his or her reputation. The value placed on any particular factor is an outcome of the interplay between interested persons and groups. There may be contextual factors relating to the working environment. For example, in scanning available jobs, the customer’s name sometimes signifies that the customer expects exceptional quality. To achieve such quality, the scheduler may need to allocate the job to a particular machine with an especially good operator. As operators find producing work of very high quality, stressful, the scheduler tries not to overload an operator with exacting work. While AI approaches can, to some extent, deal with context, they suffer from an inability to decipher qualitative differences in meaning without having them being explicitly stated (Papantonopoulos, 1990). This requires teasing out all possible contexts and associated meanings for circumstances that may have not yet arisen, and, then formulation of appropriate rules to place in a knowledge-base. Under special circumstances (e.g., when machines break down, or the characteristics of the job are abnormal) the scheduler may make changes to the machines. Schedulers can often find creative solutions through a deep-seated understanding of the resources that are available. They may transfer parts of one machine to another. During the study, three blankets used for transferring ink on the six-color press operating with 279-mm cylinders became damaged. The scheduler, who also manages production, had to act. He replaced them with two blankets used for the two-color press and one from the single-color press. To do this, he had to have functional knowledge of the machines. Removal of the blanket for the single-color press required some deliberation. For most of the shift this press was to use a 186-mm cylinder. The scheduler estimated when a job needing a 279-mm blanket was to be loaded. Confirming that spare blankets would be available by then he decided to take the blanket. There was no problem with the two-color press as the awaiting jobs did not require them. From this brief description, we can see that the environment at Melameds is complex, though not atypically so. A quite simple process of printing becomes a difficult parallelmachine problem when the machines differ in the number of colors they can produce. Sequence-dependent set-ups compound the problem. These difficulties occur although the problem has been severely reduced. Functions of processing time, set-up time and arrival time are only considered. Performance measures are restricted to simple functions such as machine utilization and tardiness. At Melameds the factors influencing scheduling extend beyond these few. They may be qualitative, subjective, and depend on context. These make the classical

8


means for ordering jobs, such as increasing processing time and first-in first-out, less significant and arguably irrelevant. Let us now consider how the scheduler produces a schedule using the machine-loading board. He places new jobs to the left of the board. Some jobs have yet to complete the previous operation: the production of printing plates. Only critical jobs are so treated; for example, jobs with impending due dates, unusually long processing times, or those that severely disturb press allocation. He looks at the current allocation and ordering of jobs at the presses. Then, he may take some unassigned cards and slot them among cards already under the chosen press heading. He often collects jobs in the unassigned space, and orders them in a desired processing sequence. He then places the group under the desired press heading, not necessarily appended to the end of the current list. As he builds a new sequence for a press, jobs previously allocated may no longer fit the current strategy. They may be shifted back to the unassigned section of the board or be allocated to another press.

5.

INTERFACE FOR HYBRID INTELLIGENT SCHEDULING

Consider the desirable features of an interface for a hybrid-intelligent scheduling system for Melameds. The scheduling system must support the scheduler’s use of job attributes. The machine-loading board provides a good starting point. It records the order jobs will load on the presses, and it allows the scheduler to observe job attributes. The system should allow the scheduler to follow practices that he has found to produce at least acceptable schedules, especially since there is no other singularly successful scheduling technique. What is being put forward is not an ethnographic design methodology in which the current practices of the scheduler are replicated in the new system without question. A new system should discourage practices that are obviously detrimental to the production of a good schedule and to encourage new proficient practices. Nevertheless, a system interface that feels somehow familiar avoids a major hurdle, user resistance. The value offered by the technology has to be great enough for schedulers to reconfigure themselves and their work to take advantage of it (Miller, Sullivan and Tyler, 1991). It should extend users so they may restructure their view of the problem. To enhance the scheduler’s ability the system has to be flexible so the scheduler can cope with unanticipated variability (Woods and Roth, 1988). The aim therefore is to allow schedulers to build upon their current practices. Figure 2 shows the job details, but not their form, that are displayed in the Job Screens. All attributes that may influence the scheduler in making decisions must be accessible. At Melameds, there are more attributes than these on a card, and they are also present in the Job Screens. For readability of the figures in this paper, and, to ensure the reader is not overwhelmed, these additional attributes have not been included. One has to be careful to separate function from implementation. Instead of merely replicating the details displayed on the job cards, one needs to consider how the scheduler uses each attribute. The manual system lists the number of sheets in a multipart form and the number of forms in the job. He uses these to determine the processing time. Usually the scheduler calculates the approximate time to manufacture in his head. Displaying processing time is more useful, especially since for difficult and critical cases he uses a calculator to determine the processing time:

9


Tproc =

Nc × D × N p × N f

+ ts v where Nc is the number passes to transfer all the colors, D is the depth of the form, Np is the number of parts in a form Nf is the number of forms, ts is the set-up time for the press, which depends upon cylinder size, colors, width and depth and v is average machine speed of the press. In the display processing time has replaced the number of forms and the number of parts, as neither of them affects scheduling decisions. Processing time is neither an attribute of a job nor of a machine. It is a property of the entity-relationship between job and machine. Therefore, the screen shows eight job attributes and four attributes derived from entity-relationships between job and machine. The job attributes are job number, customer, due date, depth, number of colors, width, and the colors on the front and back of the bill. The job-machine attributes are cylinder size, processing time, permissible presses and chosen press. For ease of discussion both job attributes and job-machine attributes are called job attributes. The Jobs Screens display the attributes of the jobs. The values of attributes are abstract signs signifying the characteristics of the printed forms (Morris, 1946; Polanyi, 1962). The significance of attributes and patterns among attributes is highly dependent upon the scheduler’s experiential knowledge of the working environment. Let us look at scheduling as it is broadly practiced. At the time a schedule is constructed, the jobs do not have a physical presence. Schedulers only have before them signs signifying job attributes. Their disposition to respond to the signs depends upon their experiential knowledge. In creating schedules, schedulers tend to invoke mental images of jobs being processed (Eberts and Eberts, 1989). By manipulating and comparing representational images, they draw out relationships. Sometimes they resort to constructing physical symbols to represent signs in a form that helps perceptual inference. For example, in comparing jobs with different values of a particular attribute they may represent the different values as lines sketched on paper. The degree of mental imagery, and the form and extent of mental modeling, are currently at the level of conjecture. Nevertheless, it is a reasonable presumption to say that in scheduling practice, abstract signs (usually displayed as alphanumeric characters), with representational images produced in the schedulers’minds, prevail. There is not a one-to-one relationship between an attribute and its signification. Signification is revealed from the interaction of the search process (i.e., the revelation of the complete state of the set of jobs being scheduled) and experiential knowledge. This process is mentally demanding. There is much mental effort required to manipulate mental images of jobs and their attributes. The aim is to find signs that support schedulers exercising experiential knowledge and mental imagery. For the signs to be effective they need to:

1. 2. 3. 4.

10

clearly and distinctively show the job attributes; clearly display, unambiguously, the values of the attributes; support the scanning of jobs to locate those jobs having a particular attribute value; and clearly display patterns in attributes across jobs.


6.

REPRESENTATION OF ATTRIBUTES

Figure 3 shows a simplified version of the “Jobs Screen” for one press. The signs, signifying jobs, are of a composite form. Through a job’s sign a scheduler accesses all the job’s attributes. Attributes map to spatial locations within the composite sign and either alphanumeric characters or graphical shapes signify their values. Each part of the composite sign is quite distinctive. In those parts of the composite sign that have insufficient retrieval cues, when the Job Screen is in full-window display a heading reminds the scheduler of the referent attribute. A horizontal line links all the components of the sign into a unified image characterising the job. As the signs used in figure 3 are abstract, users have to learn what they denote. Also there is not a one-to-one relationship between an attribute and its signification. Signification is revealed from the interaction of the search process (that is, the revelation of the complete state of the set of jobs being scheduled) and experiential knowledge. Higgins (1994a) discusses these issues in detail. Where it is possible, graphics are chosen for displaying attributes over alphanumeric characters for three reasons. First, small graphical objects can be still be visually conspicuous (Higgins, 1994a). By using small objects to depict attributes, more jobs can reside on a single screen. The user can compare in a scan the attributes for a larger set of jobs. The screen displays six tiled windows when the scheduler wants to see the schedule for the entire shop (Figure 4). When the number of jobs in the display is maximum, data density is a maximum. By maximizing data density a greater proportion of the data to be perused by the scheduler — the attributes for all available jobs — simultaneously on a screen. Leung and Apperley (1993)formalized this as the maximization of efficiency. A second, and most important, reason for graphics is to support inferential processing. It is easier to make perceptual inferences using distance, size, spatial coincidence, and color comparisons, than it is to make logical inferences using mental arithmetic and numerical comparison (Casner, 1991; Larkin and Simon, 1987). The computer brings “visibility” to those attributes experienced schedulers use to schedule jobs on the shop floor. From the patterns in the values of the attributes, schedulers can see interactions and dependencies between jobs. Through inference they decide which factors are important for the current set of jobs. They reorder the jobs on the screen to obtain more desirable patterns. This involves changing the order within a Job Screen and moving jobs to other Job Screens, either singly or en bloc. Scanning available jobs, a scheduler has to distinguish between objects depicting different attributes. The scheduler has to also discriminate one attribute’s value from another. All objects depicting the same attribute must have some features in common. However, their values may differ. Hence there must be some features that can change. These objects therefore need to be two dimensional. This can be realized through separable “global” and “local” features. The user recognizes an attribute by its global features, while the local features display its value. Global features such as shape, color, size, closure may make objects signifying different attributes quite distinct. Users can search global features quickly and, with some qualification, the time to search is independent of search size (Arend, Muthig, and Wandmacher, 1987). Local features must vary in such a way that they do not mar global integrity. For each attribute in Figure 4 the global features are obvious despite differences in the local features. Across attributes the global features are distinctly different and therefore unambiguous. Across attributes the global features are distinctly different and therefore unambiguous. 11


A sign’s form depends on the way its referent attribute affects scheduling decisions. The choice between alphanumeric and graphical signs depends upon scheduling needs. Graphical signs may have local features with either substitutive or additive scales (Norman, 1991). An alphanumeric sign has a clear advantage over graphical representation in unambiguously showing the value of an attribute. Indeed, for customer and job-number attributes it is preferred. A job number is a candidate key that uniquely identifies the job. As it acts as a label, its character is nominal and is not the subject of quantitative comparisons. If, instead, the display for the job number was graphical, a user would have to decode the graphic to obtain its value. The ability of a user to read the characters clearly and distinctly sets the lower limit on size for alphanumeric displays. Substitutive scales are used to signify the values for a job’s depth, width and colors, and for the press choices and press allocation. These attributes only have discrete values. The depth that a job may take is constrained by the size of the printing cylinder. The circumference of the cylinder has to be an exact multiple of the depth. Five cylinder sizes are available. As changing a cylinder takes a long time, the scheduler is primarily interested in cylinder size, not depth. It is preferable to maintain the same values of these attributes between jobs as a time penalty occurs where there is difference. Within the elements signifying cylinder size and color, the values of the attributes map to spatial location. For the printing cylinder, the horizontal location of the bar shows its size, and the bar’s height expresses the depth as a fraction of the cylinder’s size; for example, forms of 93-mm depth and 62-mm depth would appear as a half and third height bars, respectively, on a 186-mm cylinder. In Figure 3 the five cylinder sizes are shown as vertical dashed lines. The local feature in the cylinder’s sign for job 16356 is an unbroken vertical line that is a third of the total height. The leftmost line represents a cylinder of 186-mm circumference. Hence, the depth of the job is 62 mm as is produced on an 186-mm cylinder. A change in depth while maintaining the cylinder size is but a minor set-up. It requires only a perforating tool to be added or removed. For colors, the horizontal location of vertical bars shows commonly occurring colors. To print the same color on the front and back of a form requires two modules. Where this occurs, the bar has twice the height. In the sign representing the press allocation, a large vertical line shows the machine to which the job has been allocated. The short, vertical lines indicate alternative machines that can produce the number of colors required. In the Jobs Screens the local features for attributes using substitutive scales are vertical lines. The horizontal location of these lines signifies value. Why? Decision making in scheduling is heavily reliant on matching jobs with the same attribute values. This representation strongly supports such matching. The scheduler scans along a vertical line within an elemental part of the composite sign. To obtain a match, the scheduler seeks vertical alignment. He/she only has to decide whether a mark is present or not. The author contends that this process has superiority over scanning vertically for a change in the shape of a component (e.g., Kleiner-Hartigan tree, star symbol plot, and Chernoff faces) (Higgins, 1994a). Where values affect the shape of the graphic, the user would have to discriminate between shapes. Mapping dimensional size to an attribute’s value is useful for the comparison of continuous variables. The number and spacing of the tick marks representing hours, the presence of a horizontal bar in the sign and the sign’s horizontal location together define the sign’s global features. Its local feature is the length of the horizontal bar signifying the processing time. An additive scale makes it easy for the human scheduler to compare processing times by scanning vertically. For example, the scheduler may seek a job to fill the time that remains before the 12


shift ends, say 2 h. To find a suitable job, The author believes that the scheduler’s eye fixates on the 2-h interval and then scans vertically. Experimental evidence for this seemingly reasonable hypothesis is being sought in the next stage of the research programme. Jobs in which the horizontal line terminates to the left can be processed in the required time. These activities do not rely on the observer accurately reading the value of the attribute but on comparison. The misreading of the relative order of values that are close together does not affect scheduling performance, as the margin of error between predicted and actual values of processing time is greater. This graphic therefore meets the requirements for effectiveness (Higgins, 1994a; 1994b). While additive scales generally suit size comparison, for cases requiring an exact numerical value, Arabic notation is clearly superior. For our example, arranging jobs so that width decreases between jobs saves time in setting up. In seeking a job to follow one with a width of 229 mm, the scheduler would scan the width of other jobs. If the scale was additive, the scheduler would find difficulty in deciding whether a width was 228 mm, which would incur no set-up penalty, or 230 mm, which would do so. Therefore, the object displays width in Arabic notation, a substitutive scale. Because the display is numeric, one would expect pattern matching to suffer as it does not support perceptual inference. A graphic image as shown in the bottom right-hand corner of Figure 3 helps alleviate this problem. Where the width increases between jobs, while the cylinder size remains the same, there is a time penalty in changing jobs. The perforating tool has to be changed. The graphic draws attention to situations in which the widths increase. For those attributes not in the plant’s database, the system allows them to annotate the record with comments. Where such factors are infrequent, the system may incorporate a simple visual cue to remind schedulers to include them in their decision making. In the Melameds example a large red dot next to the job number acts as the reminder. When a job that is displaying this dot is being considered in a scheduling strategy, the scheduler may access the reminder notes by clicking on the dot. Where many of the jobs display such a simple cue may not be useful.

7.

THE SCHEDULER’S ACTIVITY

The interface has six windows. Five of the windows are Job Screens in which the scheduler can see the values of the job attributes. Of these five, one window is for unallocated jobs, and the others display jobs queued at machines, ranked in order of processing. A Gantt chart is in the sixth window. It shows the times when jobs are planned to be processed. The primary screen allows the scheduler to examine all jobs. It consists of six, nearly equal, tiled windows as shown in Figure 4. It provides the scheduler a vantage point for assessing the state of the schedule across the shop. A hybrid intelligent scheduling system has to allow the scheduler to work in a familiar way. When scheduling manually, the scheduler at Melameds places new jobs to the left of the machine-loading board. On observing the current state of the schedule and the characteristics of available jobs he assigns and reassigns jobs. Frequently, he collects jobs in the unassigned space, and orders them in a desired processing sequence, before placing them as a group under the desired press heading. The new system allows him to act in a similar way. When the display is in the overview mode, all the jobs assigned and unassigned can be surveyed. By displaying all the job attributes that may affect a schedule in the Job Screens, the scheduler

13


may regard all dependencies and conceivable interactions. He may move jobs about in a Job Screen to obtain desired patterns. For example, to keep the same cylinder set up for a group of jobs, the scheduler arranges the jobs so a single vertical line passes through the selected cylinders. To minimize the time lost in changing colors, the scheduler would arrange the jobs in this group to minimize wash ups. The system does not only support the behavior of this scheduler. As jobs may be moved at will, the system would support the behavior of any scheduler who had previously worked with the machine-loading board. The scheduler can select groups of jobs and reorder them using computer-supplied scheduling rules. The system could offer various levels of heuristics. At the basic level, the scheduler could group jobs that require the same cylinder. This would avoid a major set up within the sequence. These jobs could then be reordered using SPT or FCFS. The most advanced heuristic, if a suitable one could be found, would allocate jobs to parallel machines and sequence them. Its applicability to Melameds depends upon two other conditions. It must accommodate one major and some minor set ups, and the number of machines being different between jobs. If the environment at Melameds is typical, which the work of McKay et al. (1989) seems to confirm, factors other than those used by OR scheduling rules dominate a scheduler’s behavior. When schedulers direct the decision-making process, they can give adequate consideration to these factors. Schedulers may find groups of jobs that they may treat as equal after considering all factors except those used in OR methods. They could then apply an appropriate scheduling rule to the jobs in such a group. As the scheduler moves the jobs on the screen, the knowledge-based adviser provides advice in a non-intrusive way. At Melameds, the factors used in deciding the order of jobs to minimize set-up time could be expressed as rules that suit constraint satisfaction. When a sequence arises that results in a set-up penalty, through the firing of production rules a graphic object appears to warn the user. The graphic object that warns of increasing width requires ten rules, as it does not appear just because the width increases between successive jobs. The adviser only displays the warning if there is a significant penalty. If, for example, the width increases while the cylinder size changes, as the change in width is immaterial, the no warning will appear. The adviser informs the human of hard constraints, through graphics. For example, for each job the permissible presses are shown as small vertical lines. As a scheduler gathers jobs to allocate to a machine, the permissible machines to which they can be moved en bloc is indicated by shading of the signs for suitable presses (these signs are on the right below the menu in each window in Fig. 4). Schedulers can therefore perceive and interpret the consequence of hard constraints at a time of their choosing. If the scheduler attempts to violate a hard constraint, for example placing a job on an incapable press, then the adviser intervenes. It disallows the move, and explains why to the scheduler.

8.

SUMMARY AND CONCLUSIONS

Human schedulers are central to the decision-making process in hybrid intelligent scheduling. They should actively participate in the process, and not merely alter schedules produced by a computer. The interaction process should allow schedulers to apply methods that they naturally use, but find hard to represent as algorithms (Kempf et al., 1991). For humans to play an active and coherent role in decision making, all readily accessible data that may affect a scheduler’s decisions should be displayed. The data should be signified by signs that help

14


reveal patterns in the data and help schedulers draw inferences about possible scheduling strategies. To allow schedulers to compare jobs without having to remember details on another screen, the proportion of all data on available jobs that a scheduler can see in a single screen should be maximized. As graphical signs provide high data density, where possible they should be applied. For signs to be effective they need to (1) clearly and distinctively show the job attributes, (2) clearly display, unambiguously, the values of the attributes, (3) support the scanning of jobs to locate those jobs having a particular attribute value (4) and clearly display patterns in attributes across jobs. It is extremely important to satisfy these criteria to avoid screens cluttered with poorly designed tiny graphical objects. The following features for the interface seem plausible from the hypotheses articulated. A graphical sign needs two dimensions to signify, unambiguously, an attribute and its value. “Global” and “local” features provide the means for separating an attribute and its value. Signs with such features need to be designed to allow schedulers to scan for patterns across attributes. Where an attribute has discrete values, using the horizontal placement of vertical lines as the local feature offers an effective means for scanning its value across jobs. Similarly, for attributes that are continuous variables, a scheduler can easily scan the length of horizontal lines. The author plans to test these features experimentally. Some questions providing a focus for the experimental design are: • How hard or easy can humans identify differences and similarities in the graphical symbols? • How does an object’s graphical form affect the human scheduler’s search strategy? • How many visually-depicted attributes can a human reasonably manage? • What is the upper bound on the number of jobs that a human can search and manipulate? It has been argued in this paper that problem complexity results in a need for humans to be at the core of the decision-making process. However, there is another, and most crucial reason. Moving the problem’s locus to the needs of persons who have to take responsibility for the planning of production is the critical issue in production scheduling. The emphasis is on “responsibility”. Ultimate responsibility lies with a person, not a machine. Persons, who are responsible for production, will not wholeheartedly follow a schedule, unless they believe that it will perform to their satisfaction.

9.

ACKNOWLEDGMENTS

The author would like to thank Romesh Nagarajah of Swinburne University of Technology for his comments on this manuscript, and Andrew Wirth and Malcolm Good of the University of Melbourne for their interest in, and encouragement of, this work.

15


10.

REFERENCES

Arend, U., Muthig, K.-P. and Wandmacher, J., 1987, Evidence for global feature superiority in menu selection by icons, Behaviour & Information Technology, 6, 411-426. Bauer, A., Bowden, R., Browne, J, Duggan, J. and Lyons, J., 1994. Shop Floor Control Systems: From Design to Implementation, 2nd ed. (Chapman & Hall). Bertin, J., 1981, Graphics and Graphic Information-Processing (De Gruyter). Casner, S. M., 1991, A Task-Analytic Approach to the Automated Design of Graphic Presentations. ACM Transactions on Graphics, 10(2). Eberts, R. E. and Eberts, C. G., 1989, Four Approaches to Human Computer Interaction. In Intelligent Interfaces: Theory, Research and Design P. A. Hancock and M. H. Chignell (eds.) (Elsevier Science Publishers). Fox, M. S., and Smith, S. F., 1984, ISIS: A knowledge-based system for factory scheduling, Expert Systems, 1, 25-49. French, S., 1982, Sequencing and Scheduling: An Introduction to the Mathematics of the Job-Shop (Ellis Horwood, Chichester). Gibson, R., and Laios, L., 1978, The Presentation of Information to the Job-Shop Scheduler. Human Factors, 20(6), 725-732. Higgins, P. G., 1992, Human-Computer Production Scheduling: Contribution to the Hybrid Automation Paradigm. In Ergonomics Of Hybrid Automated Systems - III: Proceedings of the Third International Conference on Human Aspects of Advanced Manufacturing and Hybrid Automation, Gelsenkirchen, August 26-28 1992, Germany, P. Brödner and W. Karwowski (eds.) (Elsevier Science Publishers, Amsterdam), pp. 211-216. Higgins, P. G., 1994a, Graphical features for aiding decision-making in production scheduling. In Harmony through Working Together: OZCHI 94 Proceedings, Melbourne, 28 November - 1 September 1994, S. Howard and Y. K. Leung (eds.) (Ergonomics Society of Australia), pp. 261-265. Higgins, P. G., 1994b, A graphical display to support human-computer decision-making in production scheduling. In Advances in Agile Manufacturing: Fourth International Conference on Human Aspects of Advanced Manufacturing and Hybrid Automation, July 6-8 1994, Manchester, P. T. Kidd and W. Karwowski (Eds.) (IOS Press), pp. 317-320. Hwang, S.-L., and Salvendy, G., 1983, Human supervisory performance and subjective responses in flexible manufacturing systems Proceedings of the Human Factors Society 28th Annual Meeting,(Santa Monica, Human Factors Society) pp. 664-669. Hwang, S.-L., and Salvendy, G., 1988, Operator performance and subjective response in control of flexible manufacturing systems, Work and Stress, 2, 27-39. Jackson, S. and Browne, J., 1989, An Interactive Scheduler for Production Activity Control. International Journal of Computer Integrated Manufacturing, 2(1), 2-15. Kempf, K., Le Pape, C., Smith, S. F., and Fox, B. R., 1991, Issues in the Design if AI-Based Schedulers: A Workshop Report, AI Magazine, Jan 11-5, 37-46. Larkin, J., and Simon, H., 1987, Why a diagram is (sometimes) worth 10,000 words, Cognitive Science 11, 65-99. Leung, Y., and Apperley, M. D., 1993, E3: Towards the Metrication of Graphical Techniques for Large Data Sets. In Selected papers of the third international conference on Human-Computer Interaction (EWHCI'93) Moscow, August 1993, L. J. Bass, J. Gornostaev, and C. Unger (eds.) (Springer-Verlag, Berlin). McKay, K. N., 1987, Conceptual Framework for Job Shop Scheduling, MASc Dissertation, Department of Management Science, University of Waterloo. McKay, K. N., Safayeni, F. R., and Buzacott, J. A., 1988, Job-Shop Scheduling Theory: What is Relevant?, Interfaces, 18(4), 84-90. McKay, K. N., Buzacott, J. A., and Safayeni, F. R., 1989, The Scheduler's Knowledge of Uncertainty: The Missing Link. In Knowledge Based Production Management Systems, J. Browne (ed.) (North-Holland, Amsterdam), 171-189. Miller, J. R., Sullivan, J. W. and Tyler, S. W., 1991, Introduction. In Intelligent User Interfaces, J. W. Sullivan and S. W. Tyler (eds.) (Addison-Wesley, Reading, Mass.), pp. 1-10. Morris, C. W., 1946, Signs Language and Behavior (G. Braziller, New York).

16

Higgins, P.G. (1996) Interaction in Hybrid Intelligent Scheduling. International Journal of Human Factors In Manufacturing, 6(3), 185-203 Nakamura, N. and Salvendy, G., 1987, Human Decision Making in Computer-Based Scheduling within a Flexible Manufacturing System: An Experimental Study. In G. Salvendy (ed.) Cognitive Engineering in the Design of Human-Computer Interaction and Expert Systems (Elsevier Science Publishers, Amsterdam) pp. 257-264. Nakamura, N. and Salvendy, G., 1994, Human Planner and Scheduler. In, Design of Work and Development of Personnel in Advanced Manufacturing, G. Salvendy and W. Karwowski (eds.) (John Wiley & Sons, New York), pp. 331-354. Norman, D. A., 1991, Cognitive Artifacts. In Designing interaction: Psychology at the human-computer interface, by J. M. Carroll (ed.) (Cambridge University Press, New York), pp. 17- 38. Papantonopoulos, S., 1990, A Decision Model for Cognitive Task Allocation, Doctoral thesis, Purdue University. Polanyi, M., 1962, Personal Knowledge: Towards a Post-Critical Philosophy (The University of Chicago Press). Rajgopal, J., and Bidanda, B., 1991, On scheduling parallel machines with two setup classes, International Journal of Production Research, 29(12), 2443-2458. Sanderson, P., 1989, The human planning and scheduling roles in advanced manufacturing systems: an emerging human factors domain, Human Factors, 31(6), 635-666. Sheridan, T. B., 1980, Theory of man-machine interaction as related to computerized automation. In Manmachine interfaces for industrial control, by E. J. Kompass and T. J. Williams (eds.) (Control Engineering, Barrington, IL), pp. 35-50. So, K. C., 1990, Some heuristics for scheduling jobs on parallel machines with setups, Management Science, 36(4), 467-475. Solberg, J. J., 1989, Production Planning and Scheduling in CIM. In G. X. Ritter (Ed.), Information Processing '89 (Elsevier Science Publishers B. V., North Holland), pp. 919-925. Tabe, T. and Salvendy, G., 1988, Toward a hybrid intelligent system for scheduling and rescheduling of FMS, International Journal of Computer Integrated Manufacturing, 1(3), 154-164. Tabe, T., Yamamuro, S., and Salvendy, G., 1988, An approach to knowledge elicitation in scheduling FMS: Toward a hybrid intelligent system. In Ergonomics of Hybrid Automated Systems I, by W. Karwowski, H. R. Parsaei and W. R. Wilhem (eds.) (Elsevier Science Publishers, Amsterdam), pp. 259-266. Tang, C. S., 1990, Scheduling batches on parallel machines with major and minor set-ups European Journal of Operational Research 46, 28-37. Wittrock, R. J., 1990, Scheduling Parallel Machines with major and minor setup times, The International Journal of Flexible Manufacturing Systems, 2, 329-341. Woods, D. D, and Roth, E. M., 1988, Cognitive Systems Engineering. In Handbook of Human-Computer Interaction, M. Helander (ed.) (Elsevier Science Publishers, Amsterdam), pp. 3-43.

17


TABLE 1. The relationship between number of colors and the number of parallel machines. Colors 1 2 3 4 5 6

18

Machines 4 3 2 2 1 1

Higgins, P.G. (1996) Interaction in Hybrid Intelligent Scheduling. International Journal of Human Factors In Manufacturing, 6(3), 185-203 Scheduling Rules

Knowledge-Based Adviser

GANTT CHART

JOBS SCREENS

Timing at resources Performance prediction

Unassigned Sequence Job attributes Machine 1 Sequence Job attributes

Machine n Sequence Job attributes

HUMAN DECISION MAKING Context Setting Pattern Recognition

Figure 1 Interface elements and their location in a hybrid intelligent scheduling system

19

blue black red blue blue brown blue

black black grey

black

Figure 2 Job Details observable in the Job Screens

20

blue

229 345 244 49 305 149 246 229 84 254

Press chosen

2 1 2 2 1 1 3

Permissible Presses

279 186 186 216 186 186 186

Width

Front of Bill Back of Bill

279 186 186 216 62 30/6 186 186

6/5 25/5 30/4 27/5

Processing Time

Colours

Gillette TDCar Triad Jarvis Triad Option KPMG

Cylinder

16504 16667 16498 16748 16356 16537 16656

Depth

Due Date


2,3,4 1,2,3,4 2,3,4 2,3,4 1,2,3,4 1,2,3,4 3,4

3 3 3 3 3 3 3


Figure 3 Jobs Screen for Press 3: Jobs and their characteristics A partial display of objects representing job attributes

21


Figure 4 Overview of the Shop during schedule construction showing the Job Screens for the four presses and the unallocated jobs.

22

Interaction in hybrid intelligent scheduling - CiteSeerX

Interaction in hybrid intelligent scheduling - CiteSeerX

Suggest Documents

Intelligent Scheduling of Hybrid and Electric Vehicle ... - CiteSeerX

Job Shop Scheduling: Hybrid Intelligent Human-Computer ... - CiteSeerX

Multiprocessor Scheduling Using Hybrid Particle Swarm ... - CiteSeerX

The Hybrid Flow Shop Scheduling Problem - CiteSeerX

Hybrid Human-Computer Production Scheduling - CiteSeerX

Multiprocessor Scheduling Using Hybrid Particle Swarm ... - CiteSeerX

Cost-Efficient Scheduling of Elastic Processes in Hybrid ... - CiteSeerX

Intelligent strategies for scheduling in distributed

An Intelligent Preventive Maintenance Scheduling in ...

Natural, Intelligent and Effective Interaction in Multimodal ... - CiteSeerX

Optimal scheduling in the hybrid-cloud.

Optimal scheduling in the hybrid-cloud.

Scheduling in hybrid clouds - IC-Unicamp

Energy-Saving Traffic Scheduling in Hybrid

Intelligent Energy Management in Hybrid Electric ...

Hybrid Intelligent Systems in Ubiquitous Computing

Order to Disorder Transitions in Hybrid Intelligent

Intelligent Interaction: dynamic trends in today's logic

Human-Friendly Interaction in Intelligent Sweet Home

Intelligent Interaction: dynamic trends in today's logic

Geo-intelligent Traffic Scheduling For Multi-Homed On ... - CiteSeerX

HYBRID NEURAL INTELLIGENT SYSTEM TO

INTELLIGENT INTERACTION FOR SMART WORK

Intelligent scheduling: Past, present and future - Intelligent Systems ...