informatics

This article was downloaded by: [174.110.47.163] On: 09 February 2018, At: 06:53 Publisher: Institute for Operations Research and the Management Sciences (INFORMS) INFORMS is located in Maryland, USA

Interfaces

Publication details, including instructions for authors and subscription information: http://pubsonline.informs.org

Optimal Allocation of Students to Naval Nuclear-Power Training Units Michael R. Miller, Robert J. Alexander, Vincent A. Arbige, Robert F. Dell, Steven R. Kremer, Brian P. McClune, Jane E. Oppenlander, Joshua P. Tomlin

To cite this article: Michael R. Miller, Robert J. Alexander, Vincent A. Arbige, Robert F. Dell, Steven R. Kremer, Brian P. McClune, Jane E. Oppenlander, Joshua P. Tomlin (2017) Optimal Allocation of Students to Naval Nuclear-Power Training Units. Interfaces 47(4):320-335. https://doi.org/10.1287/inte.2017.0905

Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher approval, unless otherwise noted. For more information, contact [email protected].

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or support of claims made of that product, publication, or service.

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, management science, and analytics. For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org

INTERFACES Vol. 47, No. 4, July–August 2017, pp. 320–335

http://pubsonline.informs.org/journal/inte/ ISSN 0092-2102 (print), ISSN 1526-551X (online)

Optimal Allocation of Students to Naval Nuclear-Power Training Units Michael R. Miller,a Robert J. Alexander,a Vincent A. Arbige,a Robert F. Dell,b Steven R. Kremer,a Brian P. McClune,a Jane E. Oppenlander,c Joshua P. Tomlina a Naval Nuclear Laboratory, Kesselring Site, Schenectady, New York 12301; b Operations Research Department, Naval Postgraduate School, Monterey, California 93943; c School of Business, Clarkson University, Schenectady, New York 12308 Contact: [email protected] (MRM); [email protected] (RJA); [email protected] (VAA); [email protected] (RFD); [email protected] (SRK); [email protected] (BPM); [email protected], http://orcid.org/0000-0001-8778-6461 (JEO); [email protected] (JPT)

Received: November 12, 2015 Revised: July 11, 2016; December 1, 2016 Accepted: March 24, 2017 Published Online in Articles in Advance: June 27, 2017

https://doi.org/10.1287/inte.2017.0905

Copyright: This article was written and prepared by U.S. government employee(s) on official time and is therefore in the public domain.

Abstract. The U.S. Navy operates an impressive fleet of nuclear-powered submarines and aircraft carriers and has safely operated its nuclear fleet for more than 60 years, while steam- ing over 154 million miles. Rigorous training has been key to maintaining such an impres- sive record. The U.S. Naval Nuclear Propulsion Training Program develops, certifies, and delivers the nuclear-operator qualification training for enlisted and officer personnel oper- ating its nuclear fleet. This training finishes at one of four nuclear-power training units (NPTUs), operates under a complex set of hard and soft constraints, varies depending on the type of student, and requires significant personnel and equipment resources. We devel- oped and implemented a mixed-integer linear program (MILP) that prescribes how many students of each type to allocate to each NPTU at the start of each class (a group of stu- dents who train together) and how allocated students complete NPTU training. The use of MILP has improved student allocation by an estimated eight percent and led to significantly improved use of both NPTU personnel and equipment resources. In this paper, we describe this unique optimization application, the MILP formulation, its path to adoption, its user interface, and impacts from its development and use over the past three years.

History: This paper was refereed. Funding: The submitted manuscript has been authored by contractor of the U.S. Government [Contract

DE-NR-0000031]. Accordingly, the U.S. Government retains a non-exclusive, royalty-free license to publish or reproduce the published form of this contribution, or allow others to do so, for U.S. Government purposes.

Keywords: military • personnel: programming • integer • applications: education systems • planning: decision analysis • applications

Nuclear-powered submarines and aircraft carriers (Fig- ure 1) are key elements for the defense of the United States and for the maintenance of free and open com- merce across the world’s oceans (Department of the Navy and Department of Energy 2014). These vessels are staffed by highly trained enlisted and officer per- sonnel who operate and maintain the power-generation and propulsion systems capable of extended unsup- ported operations. The U.S. Naval Nuclear Propulsion Training Program develops, certifies, and delivers the nuclear-operator qualification training for enlisted and officer personnel who operate its nuclear fleet. This training finishes at one of four Nuclear Power Train- ing Units (NPTUs). This paper describes the benefits achieved by using a MILP to prescribe the number of students of different types to allocate to each NPTU at

the start of each class and the activity sequence for allo- cated students to complete NPTU training.

Certifying Nuclear Operators Certification as a naval nuclear operator requires rig- orous training that lasts at least one year for each of five student types referred to by the name of the certi- fication: electrician’s mate; machinist’s mate; electron- ics technician; engineering laboratory technician; and engineering officer of the watch. Each student type completes a unique training track consisting of knowl- edge and hands-on requirements. A student is cer- tified (i.e., qualified) to operate a specific area of a naval nuclear-propulsion plant only after demonstrat- ing mastery of propulsion plant equipment.

320

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs Interfaces, 2017, vol. 47, no. 4, pp. 320–335 321

Figure 1. The Nuclear-Powered Aircraft Carrier USS JOHN C STENNIS (CVN 74), with Destroyer Escort (to Left), and the Nuclear-Powered Submarine USS SEAWOLF (SSN 21) Operates on Deployment in the Pacific Ocean

Notes. Operating the nuclear power plants on these ships requires highly trained enlisted and officer personnel. (Photo from http://navy.mil.)

Depending on the student type, students attend one or more schools prior to beginning NPTU training. While required schools vary by student type, all stu- dent types must satisfactorily complete a six-month program (i.e., nuclear-power school), consisting pri- marily of classroom instruction, prior to six additional months of NPTU training. Ideally, upon completion of this classroom instruction, students immediately begin training at one of four NPTUs at one of two training sites; each site has two units. During NPTU training, students engage in a mix of classroom, simulator, and hands-on training. Delays in starting NPTU training often occur due to limited resources; this produces a backlog of students waiting to begin NPTU training. Navy leadership carefully monitors this backlog.

Each NPTU is a self-contained training facility com- posed of a nuclear reactor, simulators, classrooms, staff instructors, and other training assets. Each NPTU class consists of a group of students who train together, which is designated by a sequential number based on the fiscal year. For example, class 1501 corresponds to the first class started in fiscal year 2015. The starting week of each NPTU class is known, and all plants start classes on the same day; therefore, each plant runs

classes with the same class number. With rare excep- tions, a new class starts every eight weeks and three classes normally train simultaneously at each NPTU. Students are assigned (i.e., allocated) to a NPTU class and train together as a class.

NPTU training consists of a classroom phase (seven weeks) followed by a hands-on phase (17 weeks). Hands-on activities at the NPTU are critical because they teach students to perform the tasks that are required to safely operate nuclear reactors. Much of this hands-on training (referred to as “watchstand- ing”) takes place at “watchstations” located in either a plant that contains a nuclear reactor or at a simula- tor. Training conducted outside of the plant is referred to as “off-watch” training. Although extremely real- istic, simulator training can only satisfy a fraction of the required watchstanding. A qualified staff instruc- tor must be present at each watchstation in both the plant and a simulator to ensure proper plant operation. During operations, students are able to perform watch- standing. Simulators require staff time only when oper- ating for training.

There are five types of staff instructors, each per- forming different training functions. Whereas students arrive in batches at defined intervals as part of a class,

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs 322 Interfaces, 2017, vol. 47, no. 4, pp. 320–335

staff instructors continuously flow in and out of a NPTU based on their individual military assignments. Typically, a staff instructor is assigned to a NPTU for three years consistent with the typical length of other Navy assignments. In addition to supporting hands- on training in the simulator and plant, staff instruc- tors engage in a variety of other duties, including plant operation, providing classroom instruction, and administrative tasks. The amount of time devoted to each duty depends on staffing levels and the number of students being trained. It is generally desirable to allocate as many students

as possible to a class, while satisfying a variety of con- straints and assuring the efficient use of limited staff instructors, equipment, and facilities. The number and type of students impacts the use of resources; each stu- dent type has a different set of qualification require- ments, some that are unique and others that are com- mon to other student types. Assigning the number and type of students to a training class is referred to as “student allocation” (or simply “allocation”). The train- ing capability model (TCM), a MILP, prescribes how many students of each class and type to allocate to each NPTU; prescribes weekly staff-instructor assignments; and prescribes weekly student watchstanding and off- watch training.

Historical Approach For more than 20 years prior to the adoption of the TCM, a single training analyst made student alloca- tions using an iterative process, with expert judgment

Figure 2. In the 1980s, the Number of NPTUs (Solid Line Graph) Was Eight; It Is Four Today and We Expect It to Decrease to Three After 2017, While the Number of Students Requiring Training, Which We Express as a Percentage of the Peak Number Trained in 1983 (Dotted Line Graph), Has Increased in Recent Years

P e rc

e n t st

u d e n t th

ro u g h p u t

100

80

60

40

20

0 0

1

2

3

4

5

6

7

8

1 9 5 3

1 9 5 7

1 9 6 1

1 9 6 5

1 9 6 9

1 9 7 3

1 9 7 7

1 9 8 1

1 9 8 5

1 9 8 9

1 9 9 3

1 9 9 7

2 0 0 1

2 0 0 5

2 0 0 9

2 0 1 3

2 0 1 7

2 0 2 1

N o . o f N

P T

U s

Year

Students NPTUs

applied at each iteration. Because of the importance of the allocation, a second person verified calculations to ensure potential errors were minimized. The ana- lyst used a spreadsheet application to store data and assist with calculations. Over time, this spreadsheet application grew to more than 100 worksheets, which included numerous formulas and calculations, aided by Visual Basic for Applications (VBA) code (Microsoft 2015). The analyst required days to plan a single stu- dent allocation and the iterative effort was difficult to duplicate for any what-if analyses. In addition, several simplifying assumptions were employed for staff and simulator availability.

The Need for Optimization and an Expanded Model Figure 2 shows the relationship between the num- ber of nuclear operators qualifying each year and the available number of NPTUs. In recent years, the annual number of students requiring training has stayed near historic highs, while the number of NPTUs has decreased. This has necessitated the increased use of simulators and increased staffing levels. This in turn has complicated the task of determining student alloca- tions. An improved student-allocation method capable of being used by more than a single analyst and capa- ble of rapid what-if analysis was considered essential to best utilize the few remaining NPTUs.

Literature Review Naval nuclear operator training is unique, but it shares much in common with other military training.

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs Interfaces, 2017, vol. 47, no. 4, pp. 320–335 323

Military training often involves the need to complete a sequence of qualification activities, which take place at “schools.” Each school lasts a number of weeks, some- times occurs at different locations, and is only avail- able periodically. Each military specialty typically has unique schools and schools common to other military specialties. Waiting often occurs between the end of one school and the start of another. Minimizing this waiting, or minimizing the backlog in the case of the TCM, is desirable. Grant (2000) minimizes waiting time for Marines by selecting the military occupational spe- cialty (MOS) for each graduate of a common begin- ning school. Here, each MOS has its own series of schools, the timetable of when each school class begins is given, and capacity is simply the maximum num- ber of students allowed in any class. Detar (2004) and Whaley (2001) also seek to minimize the waiting time of Marines between schools by prescribing a timetable of when each school course should start and how many students should be allocated to each class, with each capacity again simply the maximum number of stu- dents allowed in any class. Capacity constraints for the TCM are more complex because each student type impacts various training resources in different ways. There is substantial operations research literature

on military manpower planning as it relates to man- aging and growing military services. Early published work on hierarchical organizations can be found in Seal (1945) and Vajda (1947). The military services employ various models to determine recruiting, pro- motion rates, and retirement (Ginther 2006, Gibson 2007, Workman 2009). Wang (2005) provides a review of operations research applications in manpower plan- ning, mostly with a focus on military training. His review includes applications that address optimization in the areas of: cost minimization for hiring and rede- ployment, personnel promotion, recruitment, and the mix and frequency of training modes (e.g., simulators, training aids) to maintain force proficiency. In general, these military manpower planning models have little in common with the TCM.

There is also substantial operations research lit- erature on the related problem of course schedul- ing, where prescriptions assign students and instruc- tors to classes, and classes to rooms and times; examples include de Werra (1985), Bonutti et al. (2012), and their extensive reference lists. The TCM

primarily differs from these course-scheduling appli- cations because different student types train simulta- neously and impact resources in different ways.

Similar resource-allocation problems can be found in healthcare literature. Caunhye et al. (2012) and Cardoen et al. (2010) provide reviews of the literature for emergency logistics and operating room planning, respectively. Examples of the allocation of operating room capacity using mixed-integer programming can be found in Zhang et al. (2009) and Blake and Donald (2002). These applications do not include prerequisite events, as required in the naval nuclear operator train- ing environment. Integer programming and simula- tion are used in decision support systems that allo- cate medical assets during public health emergencies (Lee et al. 2009) and improving emergency department operations (Lee et al. 2015). Ernst et al. (2004) give an annotated bibliography of over 700 papers on person- nel scheduling and rostering in a variety of applica- tion areas.

TCM Formulation The TCM employs an elastic MILP to determine the number of students of each type to assign to each NPTU to comprise each training class over multiple years. Additionally, the TCM determines the number of simulator sessions each week at each NPTU. It does not explicitly plan the watchstanding sequence for each individual student. Instead, for each week, it plans the off-watch hours, plant watchstanding hours for each watch, and simulator watchstanding hours for each watch for all students of each student type and class at each NPTU. It establishes a preferred watchstanding window for each class at each NPTU with soft con- straints that limit the number of watchstanding hours occurring very early, early, and late with respect to this preferred watchstanding window (Figure 3).

The TCM MILP models a number of practices that balance the competing needs for supplying qualified operators to the fleet and providing for plant main- tenance periods and limits on staff working hours. An important objective is to minimize the number of students who must wait to receive NPTU train- ing. Consequently, the first (and primary) term of the TCM objective function, which we show in Section A.5, Equation (A.1) in the appendix, imposes a penalty for each student in the training backlog (i.e., each student

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs 324 Interfaces, 2017, vol. 47, no. 4, pp. 320–335

Figure 3. The Preferred Watchstanding Window for a Class at a NPTU Occurs Between Its 14th and 21st Training Week

Preferred weeks

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Very early weeks Early weeks Late weeks

Watchstanding weeks

All training weeks

Notes. Up to five percent of the total watchstanding requirements can occur in weeks 8 and 9 and up to 50 percent of the requirements in weeks 10–13; however, any early watchstanding incurs a penalty. Late watchstanding is allowed, with increasing penalty severity, beyond week 21.

waiting to begin training). In some situations related to plant availability, a class of students is not assigned to a NPTU. This is referred to as “skipping a class” and should be avoided if at all possible. The second term of the objective function penalizes skipping classes. The third term of the objective function limits staff time to only what is needed to provide student training. Finally, a set of elastic penalties guide a solution to numerous goals. Next, we give a summary of the primary prescrip-

tions and constraints to provide a general understand- ing of the richness of the TCM MILP. Mathematical details can be found in the appendix.

The primary TCM variables are as follows. • The integer number of each student type to start

each class at each NPTU; • The integer number of each student type waiting

for training after the start of each class; • The integer number of simulator sessions for each

simulator at each NPTU each week; • The number of hours each student type in each

class performs each watchstanding requirement in its NPTU plant each week;

• The number of hours each student type in each class performs each watchstanding requirement in its NPTU simulator each week;

• The number of hours each student type in each class at each NPTU performs off-watch training each week; and

• The number of hours each staff-instructor type at each NPTU is assigned for off-watch and simulator watch instruction each week.

Weorganizedthehardandsoftconstraintsforassign- ing students to a training class and NPTU into five groups. In the following description, we use “goal” for a soft constraint (a constraint that can be violated at a cost), “limit” for a hard constraint, “each” for a con- straint that exists for each permitted value of an index, and “all” when summing over all permitted index val- ues. In Sections A.3 and A.5 in the appendix, we give the individual constraint sets and associated mathematical details for each group in the order presented here.

The class-composition constraint group establishes goals and limits on class size and distribution of stu- dent types; see constraints (A.2)–(A.7) in Sections A.3 and A.5 in the appendix. These constraints include:

• Bookkeeping to keep track of the backlog of each student type waiting for training after the start of each class;

• A lower goal and an upper goal for each student type in each class at each NPTU;

• A lower limit and an upper goal on all students in each class at each NPTU;

• An upper limit on all students at each NPTU across all simultaneous classes;

• An upper limit on all students in all simultaneous classes at each site; and

• An upper limit on all students of each student type in each class at each site.

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs Interfaces, 2017, vol. 47, no. 4, pp. 320–335 325

The student-training constraint group establishes goals and limits on watchstanding in each NPTU simulator and plant; see Constraints (A.8)–(A.14) in Sections A.3 and A.5 in the appendix. These constraints include:

• A lower limit on required training hours for each student type in each class at each NPTU and each watchstation;

• An upper limit on the total training hours for each student type in each class at each NPTU for each watch in each week;

• An upper goal on total training hours for each student type in each class at each NPTU each week;

• An upper limit on total simulator and off-watch hours each week for all simultaneous classes for each student type at each NPTU; the upper limit is set by TCM decisions on staff-instructor assignments;

• An upper limit on total simulator watchstanding for each watch across all weeks for each student type in each class at each NPTU;

• An upper limit on simulation watchstanding for each watch across all classes and all student types for each week at each NPTU; and

• An upper limit on each NPTU plant’s watch hours each week for all simultaneous classes and all student types. The staff-instructor work constraint group estab-

lishes goals and limits on the assignment of staff hours; see constraints (A.15)–(A.17) in Sections A.3 and A.5 in the appendix. These constraints include:

• An upper goal on staff-instructor hours available for simulator and off-watch instruction for each staff type at each NPTU each week; both weekly control lim- its and cumulative sustained limits on staff-instructor availability are set;

• A lower limit on the number of staff hours re- quired for each simulator session watch at each NPTU each week; the lower limit is set by a TCM decision on the number of simulator sessions; and

• A lower goal and an upper goal on simulator ses- sions for each simulator at each NPTU each week. The watch placement constraint group establishes

goals and limits on the pace of student training rela- tive to the preferred watchstanding window; see con- straints (A.18)–(A.21) in Sections A.3 and A.5 in the appendix. These constraints include:

• An upper goal on the percentage of watchstand- ing to be completed prior to a very early week (and an

early week) for each student type in each class at each NPTU and each watch; and

• An upper limit on the percentage of student off- watch hours relative to watchstanding hours for each student type in each class at each NPTU during early and late weeks.

Finally, the persistent (Brown et al. 1997) con- straint group establishes goals and limits on adher- ence to a desired partial solution; see constraints (A.22) and (A.23) in Sections A.3 and A.5 in the appendix.

TCM Implementation Features Personnel impacted by TCM prescriptions drive many of the features of TCM and its objective function. For any student allocation, the throughput of students (and the expedient reduction of any student backlog) are of primary concern, but secondary considerations, such as the efficient use of staff and plant training resources, are also considered. The TCM reflects this tiered priority structure with different penalty values for its objective function terms. The implementation also includes a time-based “reverse” discount factor applied to encourage students to complete training ear- lier rather than later in the planning horizon.

Implementers and approvers of the TCM prescrip- tions expect that small perturbations in inputs to the TCM (e.g., a small adjustment in staff-instructor hours available for weeks during a class) will yield no or at most only small changes in the TCM student-allocation prescriptions. We address this expectation by imple- menting persistence (Brown et al. 1997) as a feature, allowing a previous student allocation to be referenced as a preferred target.

Additional features are driven by practical consid- erations. Although a typical planning horizon spans two to four years, 10-year student allocations are often required to evaluate the expected long-term per- formance of the training program. For such long- term instances, the TCM employs another time-based discount that ensures near-term constraint violations (excluding those for classes already in training at the time of the allocation, which we designate as “fixed” in the appendix) are penalized more than violations that occur further in the future. Solutions over longer horizons can quickly stretch both the bounds of run- time practicality and the limits of desktop computer

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs 326 Interfaces, 2017, vol. 47, no. 4, pp. 320–335

resources. To ensure the TCM provides long-term pre- scriptions on standard desktop machines, we imple- ment a solution cascade or receding-horizon solution (Brown et al. 1987, Baker and Rosenthal 1998). This method uses a rolling horizon to solve overlapping subsets of the planning horizon, thus reducing solution time. While this method has no guarantee of optimal- ity, in practice, the TCM prescriptions are face valid using a solution cascade, and TCM users now prefer this solution method. We implement the TCM using the General Algebraic

Modeling System (GAMS) (GAMS 2015a) and solve it using CPLEX (GAMS 2015b). For a typical cascade subset of the planning horizon, about 1.5 years, an instance of the TCM has approximately 500,000 con- straints and 950,000 variables (950 of them integer). The cascade subsets typically overlap by 0.5 years. Solution time per cascade is approximately five minutes using a Windows 7 workstation with two 3.33 GHz Intel Xeon X5680 processors and 8 GB RAM. Typical TCM plan- ning horizons of 2.5 years, for example, require two cascade iterations and take approximately 10 minutes to solve, while long-term student allocations of 10 years typically require about 50 minutes.

TCM users find solution times without the use of a rolling horizon undesirable. For example, we recently conducted some experiments with real instances that have a planning horizon of 3.5 years. Using the pre- ferred method of three cascade iterations, these took between 10 and 15 minutes to solve. Solving these instances without a rolling horizon took between 30 and 45 minutes. An examination of their respec- tive solutions showed the results obtained using both methods were almost identical; the distributions of stu- dents to classes and weekly instructor staff workload varied only slightly. Empirical results such as these have led TCM users to almost exclusively use solution cascades. Despite this strong preference, we maintain an interface option for TCM users to easily disable solu- tion cascades or adjust the subset of the planning hori- zon considered for each cascade subset.

For improved solution time, TCM users are also often willing to accept a solution that is only guaran- teed to be within approximately one percent of opti- mality. With such a permitted gap, staff assigned hours (the third term of the objective function) may exceed

the number required for a given allocation. This cos- metic annoyance is corrected by solving a revised MILP that fixes the student allocation, thereby fixing the first two terms of the objective function, and minimizes the third term of the objective function.

TCM Testing Before Adoption Several management layers were needed to approve the adoption of the TCM; therefore, we designed a rigorous test program. At the time of the TCM devel- opment, decision makers had no approved and uni- versally applied criteria when evaluating allocations. Management judged new student allocations primar- ily based on the backlog and how they compared with historic allocations. Given this history, the first phase in the test program was to engage a broad range of stakeholders to develop objectives, criteria, and met- rics for allocations using value-focused thinking tech- niques (Keeney 1994). Ewing et al. (2006) and Parnell and West (2011) provide examples of applying value- focused thinking. This resulted in the following overar- ching objective statement: The allocation process seeks to maximize student throughput with on-time training completion while efficiently utilizing staff and facili- ties. The characteristics of a good allocation were iden- tified as equity in class assignments across NPTUs, on time completion, and full utilization of training resources. We then defined a set of 11 questions reflect- ing these characteristics for use by training experts in evaluating an allocation.

Simultaneously, training experts established 13 benchmark test cases representing a range of routine and abnormal scenarios. For each test case, several training experts were asked to independently evalu- ate TCM results using the 11 established questions, each rated on a four-point Likert scale (i.e., unsatisfac- tory, marginal, good, excellent). Evaluations, including open-ended comments, were submitted to a database created to enable long-term collection and analysis of TCM results. Once all test cases had been evaluated, statistics were computed quantifying the acceptability of the allocation and the associated inter-rater reliabil- ity using the method from Fleiss (1971).

Each testing round concluded with a meeting of the evaluators and model developers to discuss the results and agree on model refinements, if needed. For

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs Interfaces, 2017, vol. 47, no. 4, pp. 320–335 327

example, one round resulted in changes to the staff- instructor workload model. The evaluators found it unacceptable to expect sustained periods of high staff- instructor workload, even in theoretical cases of low instructor staffing with high demand for qualified stu- dents. A new TCM constraint was implemented to con- trol sustained staff-instructor workload. This feature allows a periodic surge of high workload if it is bal- anced by a period of lower workload. In addition to benchmark testing, a series of retro-

spective tests were performed using the most recent two years of data. In the first test, the TCM prescribed student allocations, given actual plant availability and staffing levels. In the second test, actual student allo- cations were also fixed, with the objective of solving for staff workload. The TCM has elastic constraints that allow minor violations, and we experimented exten- sively with penalties for these violations to ensure that all penalties were scaled, such that each was mean- ingful, and to capture the trade-off between exces- sive staff workload and having a backlog of students waiting for training. Following this experimentation, default penalty values were established where the stu- dent throughput was less than a five percent difference from historical values and staff workload was within acceptable ranges.

Near the end of development, the TCM was run in parallel with the legacy model for several allocations. In the parallel operations, the primary training plan- ner found the results of the TCM acceptable. In addi- tion, the TCM provided insights that were previously unavailable. These insights, coupled with the results from the benchmark and retrospective testing, were reviewed with training management who requested immediate adoption of the TCM.

Interface Design One principle analyst and several assistants are respon- sible for using the TCM to produce student allocations. Preparing TCM input requires considerable knowledge of student-training database systems and the TCM is run many times each week. A broad range of deci- sion makers rely on its prescriptions to both deter- mine the student allocation and to plan (e.g., staff- instructor schedules and plant maintenance periods). In addition, decision makers frequently request what-if analyses representing different training scenarios. All

these decision makers require graphical displays of the prescriptions.

A Microsoft Excel VBA application serves as the TCM interface. It contains all TCM documentation and input spreadsheets, serves to obtain input data from several independent databases, calls GAMS, produces all the TCM output reports, and displays all the graph- ics. The TCM produces 15 standard graphs; Figures 4–9 show examples. The TCM replicated the basic look and feel of all legacy graphics, while also providing new visualizations to display information not previously available in the legacy application.

Results and Impact Over three years, the insights gained from the use of the TCM have increased the number of students trained by an estimated eight percent (when compared with the legacy model). This improvement stems pri- marily from a holistic understanding of how student training is impacted by the interrelationships between plant availability, staffing, facility availability, and sim- ulator utilization. The ability to rapidly conduct what- if analysis that explicitly considers these interrelation- ships has led to these new insights; as a result, decision makers have altered their allocation decisions to better balance available resources.

The key to communicating the TCM prescriptions is effective visual displays. The primary display for

Figure 4. The Student Backlog (Number and Type Represented by the Different Shades in the Bar Graph) as a Function of Time (the Horizontal Axis) Provides Executive-Level Decision Makers with Key Information Concerning the Flow of Students through the Training Program, and Forms the Basis for Comparison of What-If Scenarios Involving the Allocation of Resources

450

400

350

300

250

200

S tu

d e n ts

Time (weeks)

Student backlog

150

100

50

0

Student type 1 Student type 2

Student type 3 Student type 4

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs 328 Interfaces, 2017, vol. 47, no. 4, pp. 320–335

Figure 5. Student Backlog Is Presented for Three Cases Based on the Placement of an Extended Maintenance Period, Which Would Result in Skipping Several Classes of Students at One NPTU

0

100

200

300

400

500

600

700

800

900

2018 2019 2020 2021 2022 2023

S tu

d e n t b a ck

lo g

Year

Impact of maintenance alternatives

Baseline

Alternate 1

Alternate 2

executive-level decision makers is the expected stu- dent backlog. Figure 4 shows an example of the pro- jected backlog of four student types waiting for NPTU training over time. This figure shows that initially the student backlog is large because of resource con- straints (i.e., qualified staff instructors, plant availabil- ity, or facilities). As student-training resources become available with each succeeding class, the backlog of students reduces. Subsequently, resources are again constrained resulting in a rise in the backlog. These fluctuations continue in response to changing resource profiles. Figure 5 shows an example of a what-if scenario.

The scenario is to determine the impact on student training based on the placement of a required mainte- nance period, which will make the plant unavailable

Figure 6. For Each Staff-Instructor Type, Workload Is Controlled Within the Desired Level, Sustained Limit, and Control Limit by Varying the Number and Type of Students Allocated to Each Class

Time (week)

Workload per staff instructor

Fixed staff work

W o rk

lo a d Variable staff work

Desired limit

Control limit Sustained limit

Allocation Window

for student training for an extended period and result in several skipped classes. The baseline indicates the current schedule for the maintenance period; alter- nate 1 places the maintenance period in the schedule one year later; alternate 2 places it two years later. Schedules for the other plants are unaffected and all other resources are fixed; the only resource affected is the availability of the plant at which the maintenance is taking place. The comparison clearly shows that the baseline generates the fewest number of backlogged students; however, if the maintenance is moved from the current placement, alternate 2 would be the pre- ferred option.

The TCM prescribes weekly staff-instructor variable workload across the planning horizon for each of the five types of staff. Figure 6 shows a staff-instructor workload display for one instructor type. It displays the fixed workload (duties required to keep the plants operational) and the variable workload for simulator

Figure 7. Each Bar Represents the Total Number (and with Shading the Actual Number) of Students Allowed to Simultaneously Train at the Start of Class at a NPTU

Facilities utilization

U til

iz a tio

n

Time (weeks)

Note. The student limits are a proxy for facility capacities, such as classrooms, study cubicles, and computer stations.

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs Interfaces, 2017, vol. 47, no. 4, pp. 320–335 329

Figure 8. TCM Output Provides a Graphical Representation of Plant Availability in Conjunction with the Watchstanding Timeline Associated with Each Class

Notes. The figure presented is simplified and truncated for readability. Plant availability, student allocations, and watchstanding preferences by class are shown. A box in each plant’s row indicates the week when a class completes watchstanding. For example, at Plant 1, Class 2 completes watchstanding in Week 65.

watchstanding and off-watch training as prescribed by the TCM. Figure 6 also displays the desired staff work- load level, the maximum sustained workload limit, and the maximum control limit—a surge limit that must be offset by reduced hours a few weeks prior to or after the surge. Insights provided by the TCM into the details of

staff-instructor workload for each staff type yielded one of the greatest early benefits from adopting TCM, in part because the legacy model did not explic- itly consider instructor workload. By using TCM, it became clear that some staff types were working at capacity (and thereby preventing additional student allocations), while other instructor types had hours available. Equipped with this new insight (and verifi- cation by staff instructors), changes were implemented to better balance workload among staff-instructor types and thereby increase student allocations. These changes included modifications of work assignments and adjustments in the number and type of instructors assigned to the NPTUs.

Optimally planning staff-instructor workload also allowed increased student allocations when part of training for a class coincided with a plant shutdown. Before the TCM, a simple rule of thumb dictated that only a few or no students would be allocated to a plant

when its availability prevented watchstanding comple- tion before a standard number of weeks. Using the TCM, it became apparent that this rule of thumb was too restrictive. The TCM showed how to maintain near- normal student allocations by using increased staff- instructor hours during the shutdown for off-watch and simulator training. Sometimes this needs to be coupled with a modest extension beyond a normal training deadline, with the TCM providing the details used to obtain permission for such an exception to nor- mal operations.

The TCM also ensures that student loading for each class fits within the capacity limits of the training facil- ities. This includes limits for individual classes (e.g., classroom number and size, study cubicles, computer stations), and cumulative NPTU and site-loading lim- its for all classes projected to be at a facility at one time. TCM does this by limiting the number of students in each NPTU and each site for all simultaneous classes. Figure 7 shows an example of student capacity of the training facilities over time.

Figure 8 shows a simplified summary of the TCM results for a student allocation at the four NPTUs; these results replicate much of a legacy report used by man- agement. The timeline shows how the allocation fits within each plant’s operating schedule, with no plant watchstanding occurring during maintenance periods.

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs 330 Interfaces, 2017, vol. 47, no. 4, pp. 320–335

Figure 9. TCM Provides Detailed Information for Each Plant and Class (with Only Plant 1 Shown for Readability)

Notes. For example, in Week 66, Class 2 has one simulation session and Class 3 has 20. The last row reports the percentage of plant watch- standing completion for each class. For example, in Week 64, Class 2 has 92 percent of its watchstanding complete and Class 3 has four percent.

The relationship between the preferred watchstanding periods for the three classes appears in the top por- tion of the timeline. The timeline view includes the size and composition of each class (represented in Figures 8 and 9 by the “# Students” in each of the class boxes), simulator utilization, and class completion. Figure 9 magnifies the timeline to show detailed information for each class. The timeline view is particularly useful to those making short-term training decisions. In addition to the what-if scenario presented in

Figure 5 (i.e., variation in the placement of planned maintenance periods), other typical scenarios involve changes to the number of staff instructors, additional facilities to expand student capacity, and the type and quantity of simulation equipment that could be inte- grated into the training program. In each of these what-if scenarios, the key element is the number of students that can be trained and the cost of training them, including personnel, facilities, and (or) equip- ment costs, as compared to the required number of trained students that must be transferred to the fleet to maintain the desired staffing levels onboard sub- marines and aircraft carriers. These what-if scenarios have driven changes to the number of assigned instruc- tor staff, placement of maintenance periods, and strate- gic decisions concerning future investments in equip- ment and recapitalization.

The use of TCM, with its ability to conduct mul- tiple what-if scenarios that produce accurate and repeatable results in the current resource-constrained environment, has enabled both tactical and strategic decisions to ensure fleet staffing needs are continually

met to satisfy U.S. Navy operational and strategic requirements.

Conclusions The TCM provides an optimal use of the key resources needed to qualify naval nuclear operators in a chal- lenging operational and budgetary environment. The TCM has helped increase the number of students trained by showing how to better employ these key resources, especially staff instructors. It eliminated the long-standing dependence on the few (i.e., only one or two) experts who can perform capacity analysis. Four analysts now routinely use the TCM.

The TCM’s prescriptions were quickly found to be superior to the legacy model. Today, decision makers frequently request a broad range of what-if analyses that rely on using the TCM.

Acknowledgments The authors thank Martin Andrew, Fred Lanou, and Paul Zanella for their executive support throughout the devel- opment and deployment of the TCM, Scott Ciampa (Naval Nuclear Laboratory), Dan Arguello (Naval Nuclear Labo- ratory), Karina Rodriguez (Naval Nuclear Laboratory), and Anton Rowe (Naval Postgraduate School) for their techni- cal contributions, Torre Bissell (retired) for his mentoring and software development expertise, and LT Kai Seglem (USN) for his many contributions to initial TCM develop- ment. Additionally, the authors thank the numerous Navy junior officers whose contributions at different stages of the project ensured a successful product delivery. Finally, the authors thank the Interfaces editors and referees for their thoughtful reviews that helped improve this article.

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs Interfaces, 2017, vol. 47, no. 4, pp. 320–335 331

Appendix We present a representative TCM formulation using

o ≥,

o ≤,

and i ≤ as shorthand for elastic constraints (i.e., constraints

that can be violated at a cost per given unit of violation), as discussed in Brown et al. (1997). The elastic constraints rep- resented by

o ≥ and

o ≤ can be violated in any period, while

the elastic constraints represented by i ≤ can only be violated

for periods corresponding to the first τ classes, whose mem- bers have already begun training and are considered “fixed.” For example, if τ equals 2, with Class 1 allowed to stand watch for weeks t ∈ {1, 2} and Class 2 weeks t ∈ {1, 2, . . . , 9}, elastic constraints represented by

i ≤ could only be violated

where c ∈ {1, 2} and t ∈ {1, 2, . . . , 9}. Let c (or alias c′) be the index for class number; then the allocation for classes c ∈ {1, 2, . . . ,τ} are fixed as starting conditions for the TCM and c ∈{τ + 1,τ + 2, . . .} are not fixed. Because these fixed classes may violate desired planning, we must have the ability to violate constraints for these fixed classes. Some of the terms in the constraints are applicable for only a single student type (referred to as rate r � r3).

In the following Sections A.1–A.6, we present the TCM for- mulation for indices, index sets, parameters, variables, objec- tive function, and constraints, and then conclude with a brief description. Beale et al. (1974) originally documented this ordering, which was adhered to for decades under the label NPS format in hundreds of published theses and papers by Naval Postgraduate School (NPS) students and faculty, with acknowledgment of the original Beale et al. reference (Brown and Dell 2007), and recently reintroduced with additional guidance by Teter et al. (2016).

A.1. TCM Indices [∼Cardinality] c, c′ class [6 per year];

d training site [2]; j simulator type [2]; p NPTU [4]; r student type or rate [7]; s staff-instructor type [5]; t week of the planning horizon [208];

w, w′ watch to stand [25].

A.2. TCM Index Sets p ∈ TSd set of NPTU p at training site d;

r ∈ RWw set of all student types r that stands watch w; s ∈ SWw set of staff s that can stand watch w;

t ∈ AWcpr set of all possible watch weeks t for class c, student type r, at NPTU p;

t ∈ EWc r set of all early watch weeks t for class c for student type r;

t ∈ FWcpr set of weeks t when class c can finish at NPTU p for student type r;

t ∈ LWcpr set of all late weeks t at the end of class c at NPTU p for student type r;

t ∈ V Wc r set of all very early watch weeks t for class c for student type r;

w ∈ AS set of all simulator watches w; w ∈ OW set of watches w satisfying off-watch require-

ments; w ∈ PT set of watches w that require a plant; w ∈ SS set of all watches w where a simulator can substi-

tute; w′∈ SBw set of watches w′ that can substitute for watch w;

w′∈ SMw set of simulator watch w′ that can be substituted for watch w.

A.3. TCM Parameters [Units] Parameters for the Objective Function pcrc r penalty for student type r student waiting for training

after the start of class c [penalty per student]; pcpc p penalty for not starting class c at NPTU p [penalty per

class]; psws w penalty for staff s standing watch w [penalty per

hour].

Parameters for Constraints (A.2)–(A.7) (The Class-Composition Group)

rollr number of student type r students waiting for training at the start of planning [students];

newc r number of newly arriving student type r for class c [students];

gcprcpr, gcprcpr lower and upper goal on the number of class c students of student type r desired at NPTU p [students];

lcpc p, gcpc p lower limit and upper goal on the number of class c students (excluding student type 3) at NPTU p [students];

lcpc p upper limit on the number of students NPTU p facilities can support during the first week class c is held [students];

lcdc d upper limit on number of students Site d facil- ities can support during the first week class c is held [students];

lcdrcdr upper limit on number student type r site d facilities can support during the first week for class c [students].

Parameters for Constraints (A.8)–(A.14) (The Student-Training Group) reqcprw hours of class c watch w training required by student

type r at plant p [hours per student]; donecprw hours of watch w already completed at the start of

planning by class c at NPTU p by student type r [hours];

nwwcrw minimum number of weeks per class c, student type r, and watch w that must contain some watch- standing (divisor used to provide an upper bound on weekly training for each watch);

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs 332 Interfaces, 2017, vol. 47, no. 4, pp. 320–335

grtr t upper goal on the number of hours student type r can work in week t [hours per student];

lcrwcrw upper limit on the number of class c watch hours allowed in simulators by student type r for watch w [hours per student];

ljpwjpw hours available for simulator j, watch w for each simulator watch session at NPTU p [hours per session];

lptwptw upper limit on hours of watch w available at NPTU p in week t [hours].

Parameters for Constraints (A.15)–(A.17) (The Staff-Instructor Group)

gpstpst upper goal on hours of staff s for assignment at NPTU p in week t [hours];

hfixpst fixed hours for staff s at NPTU p in week t [hours];

hsimjpw staff hours required for each j simulator watch w session at NPTU p [hours per session];

gjptjpt, gjptjpt lower and upper goal on the number of sim- ulator j watch sessions in NPTU p in week t [sessions].

Parameters for Constraints (A.18)–(A.21) (The Watch Placement Group) grvr upper limit on the fraction of student type r very early

training allowed [hours per hours]; grer upper limit on the fraction of student type r early train-

ing allowed [hours/hours].

Parameters for Constraints (A.22) and (A.23) (The Persistent Group)

τ maximum class for index c for which allocations Xcpr are fixed for all p, r;

fixXcpr allocation of student type r from class c fixed to train at NPTU p [students];

oldXcpr prior value of Xcpr used to guide a persistent solution [student].

A.4. TCM Variables Xcpr integer number of student type r to start class c at

NPTU p; Kc r integer number of student type r waiting for training

after start of class c; Ijpt integer number of simulator j sessions at NPTU p in

week t; Gc p binary variable with value one if class c is in session at

NPTU p and zero otherwise; Fcprtw plant watch w training hours from fixed plant opera-

tions assigned; to student type r in class c at NPTU p in week t;

Ucprtw simulator watch w hours assigned to student type r in class c at NPTU p in week t;

Vcprtw Off-watch w hours assigned to student type r in class c at NPTU p in week t;

Hpstw staff s hours assigned at watch w (includes simulator and off-watch instruction only) at NPTU p in week t.

A.5. TCM Formulation

Minimize {∑

c r pcrc r Kc r +

∑ c p

pcpc p(1−Gc p)

+ ∑

ptw, s∈SWw psws w Hpstw + elastic penalties

} (A.1)

Subject to:

Kc r �rollr |c�1 +Kc−1,r |c>1 +newc r− ∑

p Xcpr, ∀c, r, (A.2)

gcprcpr o ≤Xcpr

o ≤gcprcprGc p, ∀c, p, r, (A.3)

lcpc p Gc p i ≤ ∑ r,r3

Xcpr o ≤gcpc p, ∀c, p, (A.4)

c∑ c′�c−2

∑ r,r3

Xc′p r + c∑

c′�c−1

∑ r�r3

Xc′p r i ≤lcpc p, ∀c, p, (A.5)

c∑ c′�c−2

∑ r,r3

∑ p∈TSd

Xc′p r + c∑

c′�c−1

∑ r�r3

∑ p∈TSd

Xc′p r i ≤lcdc d,∀c, d, (A.6)∑

p∈TSd Xcpr≤lcdrcdr, ∀c, d, r, (A.7)∑

t,w′∈SBw (Fcprtw′ +Vcprtw′ +Ucprtw′)

≥reqcprwXcpr−donecprw, ∀c, p, r, w, (A.8)∑ w′∈SBw

(Fcprtw′ +Vcprtw′ +Ucprtw′)

≤ (reqcprwXcpr−donecprw)

nwwcrw , ∀c, p, r, t, w, (A.9)∑

w (Fcprtw +Vcprtw +Ucprtw)

o ≤grtr t Xcpr, ∀c, p, r, t, (A.10)∑

c,r∈RWw (Vcprtw +Ucprtw)≤

∑ s∈SWw

Hpstw, ∀p, t, w, (A.11)∑ t∈AWcpr,w′∈SMw

Ucprtw′≤lcrwcrwXcpr, ∀c, p, r, w∈SS, (A.12)∑ c |t∈AWcpr,r∈RWw

Ucprtw≤ ∑

j ljpwjpwI j p t , ∀p, t, w∈AS, (A.13)∑

c,r∈RWw Fcprtw

i ≤lptwptw, ∀p, t, w∈PT, (A.14)∑

w<PT Hpstw +hfixpst

o ≤gpstpst, ∀p, s, t, (A.15)∑

s∈SWw Hpstw≥

∑ j hsimjpwIjpt, ∀p, t, w∈AS, (A.16)

gjptjpt o ≤Ijpt

o ≤gjptjpt, ∀ j, p, t, (A.17)∑

t∈V Wc r

( Fcprtw +

∑ w′∈SBw∩AS

Ucprtw′ )

o ≤grvrreqcprwXcpr, ∀c, p, r, w<OW, (A.18)∑

t∈EWc r

( Fcprtw +

∑ w′∈SBw∩AS

Ucprtw′ )

o ≤grerreqcprwXcpr, ∀c, p, r, w<OW, (A.19)

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs Interfaces, 2017, vol. 47, no. 4, pp. 320–335 333

∑ t′≥t,w∈OW Vcprt’w∑

w∈OW reqcprw ≥

∑ w∈PT Fcpr,t−1,w∑

w∈PT reqc p r w

+

∑ w∈PT,w′∈SBw∩AS Ucpr,t−1,w′∑

w∈PT,w′∈SBw∩AS reqc p r w′ , ∀c, p, r, t∈LWcpr, (A.20)∑

w∈OW Vcprtw∑ w∈OW reqcprw

≥ ∑

w∈PT Fcpr,t,w∑ w∈PT reqcprw

+

∑ w∈PT,w′∈SBw∩AS Ucpr,t,w′∑ w∈PT,w′∈SBw∩AS reqcprw′

, ∀c, p, r, t∈FWc p r , (A.21) Xcpr �fixXcpr, ∀c≤ f c, p, r, (A.22) Xcpr

o ≥oldXcpr, ∀c, p, r, (A.23)

Xc p r ≥0 and integer, ∀c, p, r, I j p t ≥0 and integer, ∀ j, p, t, Kc r ≥0 and integer, ∀c, r,

(A.24)

Gc p ∈{0,1}, ∀c, p, (A.25) Fcprtw≥0, ∀c, p, r, t, w∈PT, Ucprtw≥0, ∀c, p, r, t, w∈AS, Vcprtw≥0, ∀c, p, r , t, w∈OW, Hpstw≥0 ∀p, s, t, w<PT.

(A.26)

The objective function (A.1) expresses the total penalty value. The first term of the objective function is the weighted penalty of the student backlog; the second term is the weighted penalty for skipped classes; the third term is the weighted penalty for staff workload, and the last penalty term includes all elastic constraint violations (Section A.6 in this appendix). Constraint set (A.2) tracks student back- log (i.e., inventory of students by student type waiting to start training after the start of a class). For the first period (c � 1), the backlog is initial backlog (rollr). The student back- log increases for any newly arriving students who cannot be trained in the current class, and decreases whenever more students may be trained than those newly arriving for the current class. Constraint set (A.3) measures deviation from desired lower and upper limits for each student type in each class at each plant, and ensures that no student of any student type is assigned to a skipped class. Constraint set (A.4) mea- sures deviation from the desired upper and lower bounds for the total number of students in a class at each NPTU (excluding student type r3). Constraint set (A.4) removes the lower bound for skipped classes to ensure feasibility. Constraint set (A.5) sets an upper bound on the number of students in all simultaneous classes (classes c − 2 to c for all student types, excluding student type r3 and classes c − 1 to c for student type r3). Similarly, constraint set (A.6) sets an upper bound for total number of students a train- ing site can support across all simultaneous classes. Con- straint set (A.7) limits the number of each student type for a class at a site. Constraint set (A.8) ensures sufficient train- ing hours are assigned for each student type at each watch station (watch station includes plant and off-watch require- ments). Constraint set (A.9) limits weekly training hours for

each watch. Constraint set (A.10) restricts total weekly stu- dent work hours. Constraint set (A.11) limits student and off- watch hours to those with assigned staff. Constraint set (A.12) restricts simulation training to be no more than a user input fraction of the total for training that can be conducted using simulation. Constraint set (A.13) restricts simulator hours to those with assigned personnel. Constraint set (A.14) restricts plant hours. Constraint set (A.15) limits staff hours available for watch and off-watch duties. Constraint set (A.16) ensures adequate personnel for each simulation session. Constraint set (A.17) restricts weekly simulator sessions. Constraint set (A.18) allows no more than a user input fraction of the total plant-based training to be completed very early. Con- straint set (A.19) allows no more than a user input fraction of the training of the total plant-based training to be com- pleted early. Constraint sets (A.20) and (A.21) ensure suffi- cient off-watch training in the last weeks. Constraint (A.22) fixes the initial student allocation for the first classes. Con- straint (A.23) measures negative deviation from a prior solu- tion. Constraints (A.24)–(A.26) declare variable types.

A.6. Elastic Constraints The TCM formulation has many elastic constraints that we have already introduced using notation shorthand. Here are the details for elastic constraint set (A.27), which controls the weekly and sustained staff-instructor workload by adding a new “elastic” variable Opst for the overtime worked at NPTU p by staff type s in week t:∑

w<PT Hpstw + hfixpst ≤gpstpst + Opst, ∀ p, s, t. (A.27)

In most cases, the addition of such an elastic variable (and its corresponding penalty term in the objective function) is all that is required to convert an elastic constraint from short- hand to more traditional notation. In this case, there are addi- tional constraints on the elastic variable. Constraint set (F.1a) limits the extent of overtime allowed at NPTU p by staff type s in week t:

Opst ≤ overpst, ∀ p, s, t, (F.1a) and constraint set (F.1b) limits the sustained overtime allowed at NPTU p by staff type s over all watch weeks for each student type r and each class c:∑

t∈AWcpr Opst ≤ overap s , ∀ c, p, r, s. (F.1b)

References Baker SF, Rosenthal RE (1998) A cascade approach for staircase linear

programs. Technical Report NPS-OR-98-004, Naval Postgradu- ate School, Monterey, CA.

Beale EML, Beare GC, Tatham PB (1974) The DOAE reinforcement and redeployment study: A case study in mathematical pro- gramming. Hammer PL, Zoutendijk G, eds. Math. Programming Theory and Practice: Proc. NATO Adv. Study Inst., Figueira da Foz, Portugal (Elsevier, New York), 417–442.

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs 334 Interfaces, 2017, vol. 47, no. 4, pp. 320–335

Blake JT, Donald J (2002) Mount Sinai Hospital uses integer program- ming to allocate operating room time. Interfaces 32(2):63–73.

Bonutti A, De Cesco F, Di Gaspero L, Schaerf A (2012) Benchmark- ing curriculum-based course timetabling: Formulations, data formats, instances, validation, visualizations, and results. Ann. Oper. Res. 194(1):59–70.

Brown GG, Dell RF (2007) Formulating integer linear programs: A rogues’ gallery. INFORMS Trans. Ed. 7(2):153–159.

Brown GG, Dell RF, Wood RK (1997) Optimization and persistence. Interfaces 27(5):15–37.

Brown GG, Graves GW, Ronen D (1987) Scheduling ocean transporta- tion of crude oil. Management Sci. 33(3):335–346.

Cardoen B, Demeulemeester E, Beliën J (2010) Operating room plan- ning and scheduling: A literature review. Eur. J. Oper. Res. 201(3):921–932.

Caunhye AM, Xiaofeng N, Pokharel S (2012) Optimization models in emergency logistics: A literature review. Socio-Econom. Planning Sci. 46(1):4–13.

de Werra D (1985) An introduction to timetabling. Eur. J. Oper. Res. 19(2):151–162.

Department of the Navy and Department of Energy (2014) The United States Naval Nuclear Propulsion Program (Department of the Navy and Department of Energy, Washington, DC).

Detar PJ (2004) Scheduling Marine Corps entry-level MOS schools. Unpublished master’s thesis, Naval Postgraduate School, Mon- terey, CA.

Ernst AT, Jiang H, Krishnamoorthy M, Owens B, Sier D (2004) An annotated bibliography of personnel scheduling and rostering. Ann. Oper. Res. 127(1):21–144.

Ewing PL Jr, Tarantino W, Parnell GS (2006) Use of decision anal- ysis in the Army Base Realignment and Closure (BRAC) 2005 military value analysis. Decision Anal. 3(1):33–49.

Fleiss JL (1971) Measuring nominal scale agreement among many raters. Psych. Bull. 76(5):378–382.

GAMS (2015a) GAMS: Cutting edge modeling. Accessed July 15, 2015, http://www.gams.com.

GAMS (2015b) CPLEX 12. Accessed July 15, 2015, http://www.gams .com.

Gibson HO (2007) The total Army competitive category optimization model: Analysis of U.S. Army officer accessions and promo- tions. Unpublished master’s thesis, Naval Postgraduate School, Monterey, CA.

Ginther TA (2006) Army Reserve enlisted aggregate flow model. Unpublished master’s thesis, Naval Postgraduate School, Monterey, CA.

Grant JM (2000) Minimizing time awaiting training for graduates of the basic school. Unpublished master’s thesis, Naval Postgradu- ate School, Monterey, CA.

Keeney RL (1994) Creativity in decision making with value-focused thinking. Sloan Management Rev. 35(4):33–41.

Lee EK, Chen C-H, Pietz F, Benecke B (2009) Modeling and opti- mizing the public-health infrastructure for emergency response. Interfaces 39(5):476–490.

Lee EK, Atallah HY, Wright MD, Post ET, Thomas IV C, Wu DT, Haley LL Jr (2015) Transforming hospital emergency depart- ment workflow and patient care. Interfaces 45(1):58–82.

Microsoft (2015) Introduction to the Visual Basic programming lan- guage. Accessed August 31, 2015, https://msdn.microsoft.com/ library/xk24xdbe(v�vs.90).aspx.

Parnell GS, West PD (2011) Systems decision process overview. Par- nell GS, Driscoll PJ, Henderson DL, eds. Decision Making in Systems Engineering and Management (John Wiley & Sons, New York), 275–295.

Seal HL (1945) The mathematics of a population composed of k stationary strata each recruited from the stratum below and supported at the lowest level by a uniform number of annual entrants. Biometrika 33(3):226–230.

Teter MD, Newman AM, Weiss M (2016) Consistent notation for presenting complex optimization models in technical writing. Surveys Oper. Res. Management Sci. 21(1):1–17.

Vajda S (1947) The stratified semi-stationary population. Biometrika 34(3–4):243–254.

Wang J (2005) A review of operations research applications in workforce planning and potential modeling of military train- ing. Report DSTO-TR-1688, DSTO Systems Sciences Laboratory, Edinburgh, South Australia, Australia.

Whaley DL (2001) Scheduling the recruiting and MOS training of enlisted Marines. Unpublished master’s thesis, Naval Postgrad- uate School, Monterey, CA.

Workman PE (2009) Optimizing security force. Unpublished master’s thesis, Naval Postgraduate School, Monterey, CA.

Zhang B, Murali P, Dessouky M, Belson D (2009) A mixed-integer programming approach for allocating operating room capacity. J. Oper. Res. Soc. 60(5):663–673.

Verification Letter Geoffrey Guido, Manager, Training Programs and Tech-

nology Naval, Training and Simulation, Naval Nuclear Lab- oratory, Kesselring Site, Schenectady, NY 12301, writes:

“This is to verify that the claims made in the manuscript ‘Optimal Allocation of Students to Naval Nuclear Power Training Units’ are accurate. This manuscript describes the development and use of an application that optimizes the number of students assigned to Naval Nuclear Power Train- ing Units. Since its adoption two years ago, we have realized a gain in student throughput of approximately eight percent and have significantly improved the deployment of our per- sonnel and equipment.”

Michael R. Miller is an engineer with the Naval Nuclear Laboratory. He spent 12 years in the Naval Nuclear Labora- tory’s Nuclear Operations Program training naval personnel on an operating reactor plant. He leads a software develop- ment group dedicated to providing analytical solutions for the Naval nuclear training community. He holds a Bachelor of Science in Mechanical Engineering from the University at Buffalo.

Robert J. Alexander works at the Naval Nuclear Lab- oratory developing analytic methods to support engineer- ing and training efforts. Robert earned his Bachelor of Sci- ence degree in chemical engineering from Brigham Young University.

Vincent A. Arbige works at the Naval Nuclear Laboratory developing analytical data models to support allocation of training resources. Vincent earned his Bachelor of Science degree in chemistry from the University of Rhode Island and has worked in the Naval Nuclear Training Program for 36 years. He spent 10 years in various training positions and then built the first resource allocation models for the training program, which he has supported for the past 26 years.

Robert F. Dell is a professor of operations research (OR) at the Naval Postgraduate School (NPS). He joined NPS as

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .

Miller et al.: Optimal Allocation of Students to Naval NPTUs Interfaces, 2017, vol. 47, no. 4, pp. 320–335 335

an OR assistant professor in 1990 and served as chairman of the OR Department from 2009 to 2015. During his tenure as chairman the department received the 2013 INFORMS Smith Prize. Professor Dell has been awarded the Barchi, Koopman, and Rist prizes for military operations research. He has also received two Department of the Army Payne Memorial Awards for Excellence in Analysis and two Depart- ment of the Navy Superior Civilian Service Awards. Profes- sor Dell is editor-in-chief of the Military Operations Research Journal.

Steven R. Kremer works at the Naval Nuclear Labora- tory developing analytic methods to support training. He is a retired Navy Captain who commanded USS ARCHERFISH (a nuclear powered submarine) and Naval Station Bremer- ton. He graduated from the United States Naval Academy with a Bachelor of Science degree in mechanical engineering and holds a master’s degree in political science from Auburn University at Montgomery.

Brian P. McClune is a scientist at the Naval Nuclear Laboratory. Brian divides his time between development of optimization software for the nuclear training community and development of high performance computing applica- tions in support of reactor design. He earned Bachelors of Sci- ence degrees in mathematics and physics and a Master’s of Science degree in computer science from Clarkson University.

Jane E. Oppenlander teaches statistics in the School of Business and the Bioethics Program at Clarkson University. She earned her PhD in engineering and administrative sys- tems from Union College. Jane recently retired after a 35-year career as a statistician at the Naval Nuclear Laboratory.

Joshua P. Tomlin works at the Naval Nuclear Laboratory programming systems to support engineering and training efforts. Along with his programming support, Joshua main- tains and develops processes for the Learning Management System. He earned his master’s degree in computer science from The College of Saint Rose.

D ow

nl oa

de d

fr om

i nf

or m

s. or

g by

[ 17

4. 11

0. 47

.1 63

] on

0 9

F eb

ru ar

y 20

18 , a

t 06

:5 3

. F or

p er

so na

l us

e on

ly , a

ll r

ig ht

s re

se rv

ed .