Hybrid Human-Computer Production Scheduling - CiteSeerX

30 downloads 6719 Views 33KB Size Report
Scheduling the work flow in job shops requires great skill. Companies that have .... Then the computer needs only to present information in its primary form.
Proceedings of ACME'93, Adelaide,November 1993, pp. 203-207

Hybrid Human-Computer Production Scheduling

Peter G. Higgins CIM Centre and School of Mechanical and Manufacturing Engineering, Swinburne University of Technology, P.O. Box 218 Hawthorn 3122, Australia (email [email protected])

Abstract This paper reports on the development of an interactive system for opportunistic real-time dynamic production scheduling of small-batch manufacture.

1.

INTRODUCTION

Scheduling the work flow in job shops requires great skill. Companies that have personnel who can produce proficient schedules are indeed fortunate. Whoever is responsible for scheduling needs to understand both the capabilities of the available machines and strategies for planning the sequence of operations. Where jobs have to use multiple machines, planning the order of operations can become rather complex. The number of combinations for sequencing operations easily extends beyond those that can be determined through exact mathematical solution. The presence of other contingencies further compounds by the problem. The scheduler may need to consider various choices of routings, the priority of jobs, the constraints imposed on the order of operations (precedence constraints), the need to accelerate (expedite) the progress of some jobs, the presence of random failures, the availability of materials, and even changes to production goals. A scheduler has a formidable task in defining a "good" schedule. Often multiple and disparate goals have to be sought. Consider the case where it is desired to minimise work-in-progress (WIP), and to maximise the utilisation of resources and service to customers (minimise lateness) [1]. Obviously, these individual goals conflict. For due dates to be routinely met resources have to be accessible without delay. For resources to be fully used, work has to be always available. This implies that there are jobs waiting, thereby conflicting with both the need for ready access and the minimisation of WIP. Some balance has to be found between the amount of WIP and the number of jobs that meet their due dates. Sometimes this balance is

expressed as a cost function, which is problematic. Furthermore the relative weights between goals tend to change over time. Jobs that are awaiting processing are placed in queues in front of machines in some order reflecting their priority. Priorities are based either on their attributes (due date, customer, etc.) or system attributes (the conditions of the resources) or a combination of both. Many formalised rules have been developed for the setting of priorities. They are sub-optimal strategies for forward planning work based on the state of the system at the time the schedule was constructed. Different rules are applied for the satisfaction of different objectives. As contingencies arise that divert the system's state from its predicted path, the schedule may drift further away from optimality. Shortcomings in the application of these rules have been the subject of various studies. To extend the traditional Operations Research approach, researchers have followed two paths. They either, study real human schedulers, to find out what heuristics they employ, or, develop human/computer interactive systems that compare to humans or computers acting alone. Interactive scheduling systems have a variety of forms. They may range, from systems in which humans develop a schedule, to, systems in which the computer decision making has primacy with humans only intervening when corrections and adjustments are required [2]. A common approach is to use a computer to build alternative schedules, based on different scheduling policies or dispatching rules. These are then compared and a final choice is made. Often future outcomes of a schedule may be viewed by using a predictor that steps through forthcoming events. While there is conjecture over some findings of these studies, there is general agreement that humans do make use of predictions about future states. There is general concurrence that the maximum number of steps ahead that can be readily assimilated is about three. For a comprehensive discussion of these studies, see the excellent and detailed review of the literature by Sanderson [2].

2.

HYBRID HUMAN-COMPUTER SCHEDULING

Persons are disinclined to trust decisions that vary from their own, if they cannot understand the methods and criteria that were in reaching them [3]. It is therefore necessary for computers used in interactive scheduling systems to respond according to the human's perception of the scheduling process, and to be consistent with the information-processing capacity and style of the operator. An interactive scheduling system also ought to draw on the special competencies humans bring to the scheduling activity. Meeting these criteria depends upon an understanding of how humans approach scheduling. The author has developed, and is now refining, a human-computer scheduler that attempts to fuse human intelligence and machine intelligence [6]. This "hybrid" scheduler is characterised by: 1. A coherent and active role for humans in the process; 2. Human-computer interaction during the build process; and 3. Cognitive congruence between the methods pursued by the human and the computer.

Humans must be actively involved in schedule construction. Only by being alert can they react to critical system events. To keep, or to hone, their skills in information processing, particularly inductive logic and pattern-recognition capabilities, exercise is required [7]. Therefore, the setting and maintenance of a schedule should occupy a significant portion of their attention and abilities [2]. The underlying properties in establishing the schedule should be the subject of examination, rather than mere cursory scanning of the schedule as presented. With the hybrid scheduler, unlike other interactive systems, the human intercedes during the construction of schedules, instead of just responding to choices proposed by the computer. They can place into context the diverse set of contingencies that arise. The interaction process must be able to cope with the ways that humans address the scheduling problem. While not interfering with the human's `normal' way of thinking, the computer system should extend the human's abilities through the application of suitable abstractions [8]. Three distinct types of problem-solving, characterised by Rasmussen are at the basis of the design of the scheduler. These are skill-based, rule-based, and knowledge-based reasoning. A person chooses a particular problem-solving technique for a situation based on experiential familiarity with the task at hand [9]. On being alerted to a change in state, the human scheduler may know immediately how to act. Then the computer needs only to present information in its primary form. For example, if the placing of a newly released job in the schedule can be made solely by reference to the job details (due dates, resource allocations, operations, operation times) and the current schedule. If the circumstances are such that the human scheduler needs to draw upon heuristic rules, the computer system then helps in the choice by applying rules to the primary data. The display then includes the outcomes of different rule choices. The greatest challenge is the making of decisions under circumstances where there are no known heuristics available to apply. As humans tend to solve problems in an opportunistic way, this requires tools that allow the human to move to different levels of abstraction in an ad hoc way. For all three decision-making strategies, the computer can help humans go beyond their limitations. For example, to be able to make scheduling decisions in response to rapidly changing events, outcomes, occurring days or weeks later, may need to be considered. This may be done through simulation. The computer may increase the outlook, beyond the confines of the local group of resources being scheduled, to upstream and downstream consequences, through marshalling information from other sources. While humans bring particular insight into the heuristic selection, they may need to be manoeuvred away from their cognitive biases that could lead to the taking up of an inappropriate strategy. Some biases are, fixation on the first approach to the exclusion of other possibilities, the predilection for using easily recalled information, allure of spurious cues, wishful thinking and illusion of correlation [10]. To compensate for these shortcomings, tools being developed for the hybrid scheduler suggest alternative strategies to that proposed by the human. The computer has also to accommodate other human failings. Humans may have neither an upto-date knowledge of system's state, nor an exact mental model of the system's functions and structure. They also suffer from deficiencies associated with time span. The outcomes of their

actions may not be understood until a long time later. They find difficulty in figuring out trends from a set of discrete states spaced widely apart in time. Where components in the system are tightly coupled, they may find difficulty in responding fast enough [2]. Furthermore, persons responsible for scheduling are frequently interrupted by other demands made by the production system. They have to be able to cope with the cognitive discontinuity, on returning to the scheduling task.

3.

THE INTERFACE

The format of the display is a salient feature of a hybrid scheduler. Properties have to be displayed in a form that allows for cognitive congruence between the methods pursued by the human and the computer. The syntax of the display must not obscure the data that is being enhanced [10]. Before any schedule is built, the human proposes strategies to which the computer advises on the potential consequences on the production system's performance. These have to be displayed in terms that are relevant and meaningful to the human schedulers. The visual display is critical. It allows the human scheduler to relate task objects to current information about jobs and machines. The manner information is displayed in alphanumerical and graphical form may affect how the human scheduler perceives the scheduling task [2]. In reviewing studies comparing human schedulers acting alone to their operation with a computerised aid using dispatching rules, Sanderson questioned their results. She wondered whether changes to their interface would have attenuated or even reversed them [2]. There is long experience in the use of tools for managing the large amount of data needed in the construction of a schedule. The most basic form is a machine planning board, which displays the operations on each resource for every job. A more sophisticated display, a Gantt chart, shows the sequence of operations and the expected utilisation times at each resource, including set-up times. It allows loading times and expected due dates for each operation to be determined. These devices are early forms of interactive systems. They retain information about the jobs and resources using a display that enables a human scheduler to recognise leading features easily. But, they are too cumbersome for rapidly changing environments. An early form of computerised Gantt chart was developed by Shackel to overcome the inability of board-based techniques to respond fast enough to the receipt of new orders that came on average every three minutes [11]. While the Gantt chart is useful for displaying a schedule that has been built, it is unable to help the schedule building process. During the interactive build of a schedule, the computer display has to reveal the computer's decision-making processes and consequences in terms that are meaningful to the human. During the cooperative decision-making process that precedes the build of the Gantt Chart, visual cues are presented to the human scheduler. Through pattern recognition, he/she is helped in choosing appropriate strategies that lead to a schedule that meets chosen criteria for good performance of the production system. A typical graphical aid is shown in figure 1.

depth

colours

processing time

width

16504

Smith

6/5

229

16667

JoeBlow

25/5

225

16664

Ajax

25/5

227

16654

Achme

25/5

225

16607

Cane

25/5

302

16665

Jones

25/5

303

16617

Wilds

25/5

295

16645

Hammer

25/5

225

16661

Rowe

25/5

225

16674

Thames

25/5

225

Figure 1 JOBS SCREEN - Jobs and their characteristics

4.

THE HYBRID SCHEDULER

A person responsible for scheduling enters a decision-making exchange with the computer system. A card is displayed for each job that has been released for production (see fig. 1). This card displays all data relating to the job that may be used in scheduling. For parameters in which pattern matching may be used in forming sequences, the data is displayed in graphical form (e.g. parameters in which set-ups times are dependent upon the previous job). Data in which the exact value is important, the use of a numeric display has the most clarity, as an encoded graphic does not have to be decoded to obtain the value of the measurement. A graphical aid can then be deployed to compare values, where it is desired for to have jobs ordered so that this variable is either monotonically increasing or decreasing. Cards are produced for all jobs being considered. On the screen, the cards are moved around so that they are placed into machine groups with the order of jobs reflecting their queue order. The human scheduler can consider all variables relating to the job. The human brings to the decision-making the relative importance of different factors. The cards can be sorted into groups so that particular parameters display desired patterns. Through the screen display, the human is assisted in skill-based and rule-based behaviour (see fig. 2). The computer presents information in a form that alerts the human of a state to which he/she acts immediately. For example, the scheduler may be alerted by a graphic signifying a high priority job. He/she can take immediate action and expedite the job. By arranging the cards so that desired patterns are formed, the computer helps the human in following rulebased behaviour. That is, it makes it easy for the human to carry out steps of a procedure (such as minimisation of set-up time, and, placing jobs in order of their processing time).

Dispatching Rules

Knowledge-Based Adviser

JOBS SCREEN

GANTT CHART

Symbolic objects

Timing at resources Performance prediction

HUMAN DECISION MAKING Context Setting Pattern Recognition

Figure 2 HYBRID SCHEDULING SYSTEM

Once the human has considered all the contingencies that may affect the ordering of jobs, they can be presented to the computer for advice on alternative ordering patterns that could be tried. These are based on standard operations research heuristics. Once the jobs have arranged in a desired order, a Gantt Chart is automatically built. This chart, which includes set-up times, is an outcome of the above interactive decision-making.

5.

CONCLUSION

Automated scheduling systems are unable to consider all factors that are used for scheduling jobs in a job-shop environment. Operations research heuristics may only use a single factor, and usually no more than two or three. All subjective factors are ignored.

Interactive human-computer scheduling allows the human to bring a broad range of factors to the decision-making process. The computer produces a Gantt Chart automatically using some heuristic, of a form often unknown. The human then amends the Chart using local knowledge. In "hybrid" human-computer scheduling, the Gantt Chart that is produced is a product of a decision-making process in which the human is actively.

6.

REFERENCES

1.

Emmons, H., Scheduling and Sequencing Algorithms, in: John A. White (Ed.), Production Handbook, 4th ed., John Wiley & Sons, New York, 1987. Sanderson, P. M., The Human Planning and Scheduling Roles in Advanced Manufacturing Systems: An Emerging Human Factors Domain, Human Factors, 31(6), 635-666, 1989. Bainbridge, L., Ironies of Automation, Automatica, 19, 775-779, 1983. Sanderson, P. M., Towards the Model Human Scheduler, International Journal of Human Factors in Manufacturing, 1991. Rasmussen, J., Information Processing and Human Machine Interaction: An Approach to Cognitive Engineering, North-Holland, New York, 1986. Tabe, T. and Salvendy, G., Toward a hybrid intelligent system for scheduling and rescheduling of FMS, International Journal of Computer Integrated Manufacturing, Vol. 1, No. 3, 1988. Sheridan, T. B., Toward a general model of supervisory control, in: T. B. Sheridan and G. Johannsen (Eds.), Monitoring Behavior and Supervisory Control, Plenum, New York, 1976. Sage, A. P., An Overview of System Design for Human Interaction, in: A. P. Sage (Ed.), System Design for Human Interaction, IEEE Press, New York, 1987. Moray, N., Dessouky, M. A., Kijowski, B. A., and Adapathya, R., Strategic Behavior, Workload and Performance in Task Scheduling, EPRL-90-06, University of Illinois at Urbana-Champaign, 1990. Goodstein, L. P., Discriminative Display for Process Operator, in: J. Rasmussen and W. B, Rouse, Human Detection and Diagnosis of System, Plenum Press, New York, 1981 Shackel, B., personal communication, 18th February 1992. Vicente, K. J. and Rasmussen, J., The Ecology of Human-Machine Systems II: Mediating "Direct Perception" in Complex Work Domains, EPRL-90-91, University of Illinois at Urbana-Champaign, 1990. Rasmussen J. and Vicente, K. J., Coping with Human Errors through System Design: Implications and Ecological Interface Design, International Journal of Man-Machine Studies, 31, 517-534, 1989.

2.

3. 4. 5. 6.

7.

8. 9.

10. 11. 12.

13.