Planning & Scheduling Benchmarks

Resource Constrained Project Scheduling

This data set, and its associated series of problems, is based upon large scale assembly but is applicable to a wide variety of problems in engineering, construction, and manufacturing that can generally be described as resource constrained project scheduling. This same data set and series of problems could have originated from a commercial construction project, a petro-chemical plant start-up, an automotive assembly line set-up, Space Shuttle reprocessing, or the assembly of heavy machine tools. I have prepared a first series of problems based upon this data set. Time permitting, I will prepare additional problems that are based upon this same data set or prepare additional data sets that can be the subject of this same series of problems.

In brief summary, this series of problems is based upon the most common questions raised by production management. The capabilities demonstrated in the solution of these problems can be used to answer the following:




The Focus

You will notice that every problem in the first series requires the construction of a schedule, and that the primary focus seems to be on producing "optimal" schedules, but taken as a whole the first series of problems represents an entirely different sort of challenge. My experience is that less than 1% of the code in a working scheduling application is devoted to schedule optimization. The remainder of the code is devoted to a diverse collection of functions required to make the application useful and user-friendly (graphical user interface, postscript reports, etc.) By similar measure, less than 5% of the time spent scheduling is making new schedules while 95% of the time is spent revising and maintaining schedules in the context of daily progress and changing assumptions. Hence, the real emphasis in the first series of problems is upon having a complete collection of schedule construction and maintenance capabilities that can be used to directly support production management. These problems exercise a scheduling engine with many of the very same challenges that it would encounter if deployed in a real production environment.

Even though the emphasis is not upon optimization, it is still necessary to produce good solutions to these problems. When a customer has a choice of systems that provide equivalent functionality, they will always choose the one that produces better schedules. However each industry and application has a different measure of quality. In the following problems, the focus is on minimal cycle time but in the actual factory setting there are a large number of other, sometimes subjective, measures of quality that must be respected. These other factors are not part of this first series of problems only to keep its scope limited and focused.

When measuring the quality of a schedule, it is important to agree not only upon the metric for comparison but also upon the standard for comparing two metric results. Traditional statistical measures can be used effectively to determine when two scheduling methods consistently produce different results but this should not be confused with what end users perceive to be a significant difference. In this, and other similar applications, a day is the smallest significant time quantum. This is because so many other aspects of production, like the start of a new assembly or the delivery of parts, are aligned on day boundaries. Two schedules that differ by a few minutes or a few hours will not be significantly different. In such settings, a schedule must be at least 1 day better in order to be significantly better, but a 1 day difference can be worth $100,000 or even $1,000,000. It will be noticeable that this first series of problems is devoid of any "interesting" constraints or situations. In this data set there are only two kinds of constraints, resource and precedence. And we explore only one kind of resource usage. At a later time I hope to publish additional data sets that explore other technical issues including interruption and preemption, placement preferences, resource modeling, and state constraints.




The Challenge

To some extent, the twelve problems given below are arranged according to progressive difficulty. Problem 1 is concerned only with precedence while Problem 2 is concerned with both precedence and resource constraints. But Problem 2 and Problem 3 are qualitatively quite different. Problem 2 can be solved with a scheduling engine that possesses no lookahead but Problem 3 requires sufficient lookahead to accurately schedule within shift boundaries. Problem 4 requires the ability to look past the shift boundaries so long as sufficient resources are available. Problem 5 requires the ability to look beyond shift boundaries and to perform scheduling within the context of some existing, but irregular pattern of resource availability. Problem 6 requires essentially the same capability but may prove to be prohibitively difficult for scheduling engines that perform purely chronological schedule construction. Problem 7 exercises a scheduling engine's ability to perform incremental schedule modifications, while Problem 8 exercises a system's ability to constrain its operations to specific windows of time. Problems 9 and 10 are similar but might prove difficult for systems that rely upon hard coded models of resource availability rather than explicit data representations. Problem 11 again looks much like Problem 4 but tests whether the scheduling engine can be constrained in its effects while still pursuing an optimization goal. And finally Problem 12 exercises the scalability of a scheduling engine and at the same time tests whether a scheduling engine can simultaneously optimize multiple objectives.

You will find that the problems in the first series require quite a variety of capabilities. It is not necessary that 100% of the work required to solve a problem be automated, but participants are expected to report on what parts of the solution process were accomplished through data manipulation and commands issued to their scheduling system. To be honest, the only thing that counts is the ability to produce good solutions to these problems in a cost-effective and timely fashion. Combinations of manual and automated methods will always be necessary to adapt to the continuously changing manufacturing environment.

Some participants will find the complete series of problems to be prohibitively difficult. Others will find them to be within their capability. In the end, the number and kind of solutions submitted will substantially influence perceptions of what constitutes the state-of-the-practice and the state-of-the-art. At the same time, the number and kind of solutions submitted will influence the level of difficulty in subsequent benchmark problems.

I recognize that most participants will be reluctant to publish their results without some knowledge of what others have accomplished. To reduce such anxiety, my partner in this endeavour, Mark Ringer, has agreed to work the first 4 problems with a simple first fit, interval based scheduling engine, with no optimization. His results are summarized in the section on results submitted by participants. I hope that all participants will report as many results as possible so that we can begin to determine what is a reasonable level of capability in addition to what is an achievable level of capability.




The Physical Problem

The physical problem, which this abstract data set represents, is the process of creating a large assembly consisting of a large number of discrete steps. Each step entails the performance of specific work documented in formal process plans. We can assume that the activity name, consisting of an assembly number followed by a step number, uniquely identifies each step and provides an unambiguous link to the formal process plans. This data set contains only the information required to schedule the work and all other information necessary for the performance of the work is found elsewhere. The duration of each step is annotated in hours and minutes and the work has been divided so that no single step requires more than 1 shift (7.5 hours) for its performance. Each step requires zero or more individuals drawn from 4 labor pools, Px, Py, Pz, and Pw, to be present for the full duration. While performing a step these individuals and their supporting equipment may occupy or exclusively reserve space within designated work-zones (Za, Zb, etc.) around the assembly.

This data set was synthesized from actual assembly problems and, based upon my experience, is perfectly realistic. However the specific names, durations, labor, zone, and precedence constraints have all been created to make this problem entirely unique, and dissimilar from any real assembly. Hence, labor is named by anonymous pools Px, Py, Pz, and Pw, work zones around the assembly are named Za, Zb, etc., and the steps are designated by a two part name consisting of an anonymous assembly number and an anonymous step number. In addition, the step numbers have been randomized so that there is no correlation between step number and precedence.




The Data File

The data set, consisting of four parts, contains all of the data required in the graduated series of problems described below. The format for the data was chosen not for its generality or expressiveness, but rather for its simplicity so that minimal effort would be required in the creation of readers and writers. And in order to further minimize the effort required to get started, we have provided C++ source code for routines that will read and write this data. In brief summary, section 1 of the data set contains the activity specifications consisting of name, duration, and resource requirements; section 2 of the data set contains the precedence relations between the activities specified in section 1; section 3 of the data set contains a feasible schedule for the performance of the activities specified in sections 1 and 2; and finally section 4 contains miscellaneous data found in the following problem descriptions. C++ source code that implements basic reading and writing routines for the data found in sections 1, 2, and 3 is provided in a separate file, courtesy of Mark Ringer.

Any participants that received prior beta copies of this data should destroy those old data sets and use only the data acquired from this source. There have been some significant changes since it was first created.



The complete set of problems for the first series along with all of the necessary data can be retrieved below.






Return to Planning & Scheduling Benchmarks Home Page.