Professional Documents
Culture Documents
3 HP Enterprise Services Uses Optimization For Resource Planning
3 HP Enterprise Services Uses Optimization For Resource Planning
Interfaces
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org
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 permissions@informs.org.
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.
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
Vol. 43, No. 2, March–April 2013, pp. 152–169
ISSN 0092-2102 (print) ISSN 1526-551X (online) http://dx.doi.org/10.1287/inte.1110.0621
© 2013 INFORMS
Resource Planning
Cipriano Santos, Tere Gonzalez
HP Labs, Palo Alto, California 94304
{cipriano.santos@hp.com, maria-teresa.gonzalez.diaz@hp.com}
Haitao Li
College of Business Administration, University of Missouri, St. Louis,
St. Louis, Missouri 63121, lihait@umsl.edu
Kay-Yut Chen
HP Labs, Palo Alto, California 94304, kay-yut.chen@hp.com
Dirk Beyer
MarketShare L.L.P., Los Angeles, California 90025, dbeyer@marketshare.com
Sundaresh Biligi
HP Enterprise Business, Bangalore 560 100, India, bilgi@hp.com
Qi Feng
McCombs School of Business, University of Texas at Austin, Austin, Texas 78712,
annabelle.feng@mccombs.utexas.edu
Ravindra Kumar
HP Enterprise Business, Bangalore 560 100, India, ravindra.kumar@hp.com
Shelen Jain
HP Labs, Palo Alto, California 94304, shailendra.jain@hp.com
Ranga Ramanujam
HP Enterprise Business, Bangalore 560 100, India, ranga.ramanujam@hp.com
Alex Zhang
HP Labs, Palo Alto, California 94304, alex.zhang@hp.com
The main responsibility of resource and delivery managers at Hewlett-Packard (HP) Enterprise Services (HPES)
involves matching resources (skilled professionals) with jobs that project opportunities require. The previous
Solution Opportunity Approval and Review (SOAR) process at HPES addressed uncertainty by producing
decentralized project staffing decisions. This often led to many last-minute subjective, sometimes costly, resource
allocation decisions. Based on our research, we developed a decision support tool for resource planning (RP)
to enhance the SOAR process. It optimizes matching professionals who have diverse delivery roles and skills
to jobs and projects across geographical locations while explicitly accounting for both demand and supply
uncertainties. It also embeds capabilities for managers to incorporate tacit human knowledge and judgment
information into the decision-making process. With its 2009 deployment in Best Shore, Bangalore operations of
HPES, the RP tool’s significant benefits include reduced service delivery costs, increased workforce utilization,
and profitability.
Key words: resource planning; workforce optimization; decision support; mathematical programming; supply
and demand uncertainty; human knowledge.
History: This paper was refereed. Published online in Articles in Advance August 24, 2012.
152
Santos et al.: HP Enterprise Services Uses Optimization for Resource Planning
Interfaces 43(2), pp. 152–169, © 2013 INFORMS 153
range of services, including application solutions, out- estimated, based on which workforce sourcing deci-
sourcing, and information technology (IT) services to sions resource managers make. This information is
its customers. Its workforce consists of HP regular updated and repeated as the process approaches the
and contingent employees and non-HP third-party BA and Ts&Cs event.
partners worldwide. Although SOAR provides workforce supply and
The objective of resource planning (RP) is to pro- demand matching, its previous RP methodology has
vide workforce resources who have the right skill for the following limitations. (1) Resource managers lack
the right job at the right time and at the right cost visibility of the project funnel because SOAR staffing
(Grinold and Marshall 1977). RP at HPES has three decisions are decentralized. A resource manager can
formidable challenges. The first is scale and complex- look only at the resource requirements and search
ity. Thousands of professionals with diverse service supply in a specific business domain, with no commu-
delivery roles and skills must be matched dynamically nication between other domains of service. (2) SOAR
to a myriad of projects and jobs in countries world- either does not consider the uncertainty of resource
wide. The second is uncertainty in both labor sup- demand or treats it in a primitive way—HPES staffed
ply and demand. Most demand information comes project opportunities with a win probability over a
from estimates of future project opportunities by the prespecified threshold. This approach often led to
HP sales force. The main sources of demand uncer- suboptimal solutions because it completely ignored
tainty relative to a project are whether HPES will win the resource demand incurred by those opportunities
it, when it will start, and what its associated labor with win probabilities that are less than the thresh-
needs are. On the supply side, the availability of work- old. (3) Because of limitations (1) and (2), many last-
ers is often uncertain because of attrition. The third minute decisions had to be made manually based
challenge is that matching resources and jobs must only on managers’ experiences and subjective discre-
consider multiple attributes. A well-defined objective tions. Such last-minute decisions were costly because
function for matching is not readily available. For HPES often had to resort to a more expensive contin-
example, suppose the manufacturing industry has a gent workforce (CW) to fill gaps. (4) Matching relied
demand for an expert-level Java programmer, and primarily on management judgment or soft matching
HPES has an employee (e1) who is an intermediate- rules, which resource or project managers implicitly
level Java programmer from the aerospace industry implemented. No unified and systematic approach to
and an employee (e2) who is an expert-level Java pro- performing such soft matching existed.
grammer from the telecom industry. Whether e1 or e2 We proposed a new RP solution to enhance the
is better suited for the job is unclear. Similarly, projects SOAR process. It allocates labor resources based on a
can have importance beyond their financial impact. hybrid approach that consists of a rigorous optimiza-
A small project for a strategic customer could be tion model and a set of business processes to augment
more important than one that generates higher prof- RP decisions with human judgment and knowledge.
its, because of the future profit opportunities. Much This paper focuses on the optimization and process
of the information relevant to the objective function aspects; supply and demand analysis (e.g., the calcu-
is tacit and exists only in managers’ minds. lation of win probabilities) is outside the scope of this
HPES employs an approval structure, the Solution discussion.
Opportunity Approval and Review (SOAR) pro-
cess, to evaluate project opportunities. SOAR consists
of four events: opportunity assessment (OA), bid Resource Planning at HPES
approval (BA), terms and conditions (Ts&Cs), and In this section, we describe the RP problem (see Fig-
scope change (SC). Various RP functionalities are ure 1). Labor demand is derived from job requirements
embedded in these SOAR events to match resource of project opportunities and attrition replacements.
Santos et al.: HP Enterprise Services Uses Optimization for Resource Planning
154 Interfaces 43(2), pp. 152–169, © 2013 INFORMS
Demand
Job fulfillment
Fulfillment/
requirements
Downloaded from informs.org by [203.200.225.151] on 18 September 2016, at 06:05 . For personal use only, all rights reserved.
allocation
Project
plan
opportunities
Job Attrition
characterization replacements
Labor demand
Transition
RP PLANS
Job Plan plan
attributes scenario
Labor Process
Hiring
nomenclature Current and Transition plan
future bench rules
Employee
characterization
Joinee Hiring Utilization/
pipeline rules demand
Resource
fulfillment
Labor supply allocation
reports
Both job requirements and labor supply capabilities to match, planning parameters (e.g., limits on num-
can be characterized by a labor nomenclature that is ber of people to hire, service levels, or hiring and
defined by job attributes such as location, industry, training lead times). Resource planning has four out-
skill type, and proficiency level. Labor supply is avail- put components: an employee allocation and fulfill-
able from the following sources: ment plan, an employee transition plan, a hiring plan,
• Current bench: the existing workforce capacity and reports about demand fulfillment and resource
(i.e., a group of employees waiting to be assigned to utilization at various aggregation levels. The alloca-
project opportunities) tion and fulfillment plan provides the project-job-
• Future bench: employees released from com- employee-time allocation based on preferences (busi-
pleted projects ness rules), which specify the priority for different
• Joinee pipeline: newly hired employees with cer- sources of capacity. The employee transition plan
tain lead time to be deployed describes employee training, promotions, and rede-
• Transition: employees to be qualified for projects ployment (i.e., any factor that changes an employee’s
or jobs through training and redeployment current status to satisfy job requirements). It identi-
• Hiring: employees to be hired fies opportunities to increase resource utilization and
Demand is stochastic in this setting. In particular, aligns employee career paths with the business needs
whether a project opportunity will be won, its start- of satisfying project requirements. Job requirements
ing time, and its workforce requirements (determined that the fulfillment plan and employee transition plan
by the scope of work) are uncertain. In addition, sup- cannot fill are filled by hiring new employees. The
ply is also stochastic. That is, the number of employ- demand fulfillment and resource utilization reports
ees available over the planning horizon is affected help resource managers to evaluate the current sce-
by attrition. They might also be engaged in ongoing nario. Executing these plans and realizing the random
projects for longer or shorter periods than expected. events update the state of the system for the next
A resource manager creates a scenario by defining planning period, reflecting the dynamics of the RP
the scope of labor supply and demand he (she) wants problem being addressed.
Santos et al.: HP Enterprise Services Uses Optimization for Resource Planning
Interfaces 43(2), pp. 152–169, © 2013 INFORMS 155
manpower and personnel planning. The general prob- optimization (Santos et al. 2009b) at the strategic level,
lem of manpower planning, as Gass (1991) defines and task scheduling with multiskilled personnel (Li
it, involves determining the number of workers and and Womer 2009a) at the execution level.
their skills that best meet an organization’s future
operational requirements. Research on planning a Modeling Approach
multiskilled workforce is abundant in various set- There are various ways to deal with optimization
tings. In the military area, applications include both under uncertainty. The analyst must choose a trade-
long-term strategic manpower planning (Gass et al. off between modeling capability and solution effi-
1988) and matching personnel with jobs at the short- ciency. We can view the addressed stochastic RP prob-
term operational level (Holder 2005, Li and Womer lem as a sequential decision problem, which we can
2009b). Applications in the service sector include shift model as a Markov decision process (MDP) (Puter-
scheduling (Burns and Koop 1987), nurse rostering man 2005). However, solving such a high-dimensional
(Kao and Queyranne 1985), and crew scheduling in MDP for thousands of projects and employees is
the airline (cf. Barnhart et al. 2002, Yu et al. 2004) challenging, because classical dynamic programming
and railway (Abbink et al. 2005) industries. The HPES (DP) (Bellman 1957) based on Bellman’s recursion is
resource-planning problem differs from those men- computationally intractable. Stochastic programming
tioned earlier in that it addresses matching workforce (SP) (Birge and Louveaux 1997) is another possibility,
supply with service-project based demand. which potentially leads to a large-scale mixed-integer
Recent studies on optimizing project staffing can be program (MIP), with its size growing rapidly as the
found for managing service projects in the context of a number of scenarios grows. Because HPES faces a
service center (Valls et al. 2009) and IT project staffing large-scale RP problem under uncertainty and must
(Li and Womer 2009a). In both settings, operational- solve it frequently, we did not choose an approach
level personnel assignment and task-sequencing that attempts to address uncertainty and random-
and scheduling decisions are simultaneously opti- ness over all periods in one model (e.g., MDP or
mized in the framework of resource-constrained SP). Instead, we used a two-phase approach that
project-scheduling problems (cf. Demeulemeester and first builds a buffer capacity of a workforce (planned
Herroelen 2002). In our RP problem, staffing deci- bench) to provide reliable estimates of the quantity
sions must be made at the tactical level. Each project of resources with certain attributes over the planning
opportunity is associated with a fairly predictable and horizon by accounting for the uncertainty of labor
staged set of jobs to be executed, and the sequence demand and supply. We then enter this buffer capac-
of required jobs for each opportunity is derived ity into an optimization model to obtain RP solutions.
in advance from the project plan. Therefore, rather Our modeling approach (see Figure 2) consists of
than considering task-sequencing decisions as in a two modules: a preprocessing module, supply and
typical scheduling problem, our RP problem treats demand consolidation (SDC), and an optimization
project opportunities at an aggregate level and opti- module built upon an MIP model. The SDC mod-
mizes resource-and job-matching decisions. It consid- ule takes information about project opportunities,
ers the richness and complexity of job and workforce employees, and various uncertainties as input. Its
attributes and business rules for resource demand main outputs include: (1) a qualification table that
matching, training, hiring, and relocation. determines jobs for which an employee is fully (i.e.,
Our work fills the gap between long-term strate- 100 percent) qualified; (2) a transition table that defines
gic workforce planning and short-term personnel the score for less than fully qualified matching and the
scheduling in a hierarchical modeling framework, as type of workforce transition (e.g., training, promotion
Gass (1991) and Pinedo (2005) suggest. In particular, and demotion, and relocation) required to fully qualify
Santos et al.: HP Enterprise Services Uses Optimization for Resource Planning
156 Interfaces 43(2), pp. 152–169, © 2013 INFORMS
+ Win probabilities
+ Buffer capacity — Planned
+ Starting-time probabilities
bench
+ Employee qualifications and trainability
+ Job attrition rates
Output
the employee for the job; and (3) a planned bench of deploying a resource to a project, making the delivery
employees who can serve as a buffer capacity to hedge cost visible during the RP process.
against demand and supply uncertainty. We enter • Workforce (WF) type identifies a resource’s
these outputs into the optimization module to con- source—regular workforce (RWF) or contingent work-
struct the MIP model, which optimizes demand ful- force (CW).
fillment by balancing the trade-off between workforce • Location identifies a source’s geographical lo-
utilization and availability. The optimization module cation.
must encode business rules that reflect manager pref- • Business domain (BD) expresses a resource’s area
erences in allocating employee resources. We discuss of expertise (e.g., manufacturing, aerospace, telecom-
these two modules next. munications).
• Location type identifies whether a resource is
Supply and Demand Consolidation (SDC)
onsite or offshore.
The SDC module includes three main components:
Building the planned bench: We designed the
mechanisms for consolidation, building a planned
planned bench as a buffer to address supply and
bench, and flexible matching.
Mechanisms for consolidation: The mechanism for demand uncertainty. The main source of demand
consolidation relies on a unified nomenclature of job uncertainty is the highly unpredictable sales process.
attributes (see Figure 3) that describe a job, resource, We cannot perfectly forecast which projects HPES will
and workforce transition (training or relocation). win. Thus, we assign a probability (i.e., the win prob-
• Skill group (SG) maps resource demand with the ability) to each project being pursued. Sales teams are
language of labor supply; examples include Java 2 required to update this information via the SOAR pro-
Enterprise Edition (J2EE), dot net, and Advanced Busi- cess. However, as Chen et al. (2004) discuss, human
ness Application Programming (ABAP). We obtain it estimates of probabilities are not always unbiased or
by identifying standard projects of a practice, standard accurate. Therefore, we developed a separate method-
jobs required by a standard project, and the main tech- ology to improve the accuracy of the win probability
nical skills required to perform a standard job. estimate by removing bias and increasing the effi-
• Job level defines a resource’s expertise level, ciency of information use through a behavioral eco-
including entry level (ENT), intermediate level (INT), nomics model that captures the salient characteris-
expert level (EXP), and master level (MAS). Given dif- tics of the human forecasting process. This work has
ferent pay grades associated with the expertise level, implications beyond the workforce matching issues
this attribute incorporates the cost of assigning or presented here, thus beyond the scope of this paper.
Santos et al.: HP Enterprise Services Uses Optimization for Resource Planning
Interfaces 43(2), pp. 152–169, © 2013 INFORMS 157
Job attributes
Qualified for Specified as
Defined as
Job WF Location
SG Location BD
level type type
Belongs to
Contains Example
Technology
Skill group— J2EE, dot net, ABAP
Job level —ENT, INT, EXP, MAS
Workforce type— RWF or CW
Skills
Location— Bangalore, Chennai, USA, Germany, UK
Business domain—Manufacturing, Aerospace, Finance
Location type—Onsite or Offshore
Figure 3: This diagram provides an example of the hierarchy of job attributes for SDC.
The main source of supply uncertainty is employee exact matching often results in poor demand fulfill-
attrition. In some countries, attrition rate can be ment and resource utilization: only about 40–50 per-
higher than 30 percent per year. The planned bench is cent of job requirements can be matched exactly by an
a probabilistic model that maps the uncertain project available worker. This motivated us to design a flex-
opportunities and the attrition process to service lev- ible matching method (FMM) (Gonzalez et al. 2009);
els. Appendix A provides details of the model. we use the analytical hierarchy process (AHP) (Saaty
Figure 4 illustrates an instance of planned bench 2000) approach to compute weights of job attributes
for a particular skill group. The probability distri- that reflect each job attribute’s importance to perform-
bution on the left shows a 60 percent confidence ing the job. FMM enables us to match employee capa-
interval of [8,000, 11,000] (i.e., there is a 60 percent bilities with job requirements at less than 100 percent.
probability that between 8,000 and 11,000 man-days We define a hierarchy of job attributes based on
of this skill group will be required). The bench to their relative importance (weights), and label each
achieve a 75 percent service level (SL) must have job attribute as either mandatory or flexible. If a
9,600 man-days. As we mentioned earlier, the HPES job-resource mismatch of a mandatory attribute is
business practice prior to implementing the RP tool present, the resource is not qualified to perform the
was to staff project opportunities with win probabili- job; if a mismatch of a flexible attribute is present,
ties higher than or equal to SL (75 percent). Note that we calculate a matching score to quantify the quality
if a firm only staffs opportunities with a win probabil- of the matching. We designed a rescaling mechanism,
ity greater than 75 percent, the bench size is approx- which adjusts the attribute weights, to account for
imately 6,200 man-days with an actual SL of only the dependency among job attributes and its impact
5 percent. Thus, the RP decisions based solely on win on the matching score (see Appendix B). For exam-
probability may lead to a poor SL. ple, Table 1 shows three candidate employees to
Flexible matching: Matching a job and an employee be matched with a job described by five attributes
who has 100 percent of that job’s requirements is without rescaling. Employee 1 (E1) matches the job
straightforward. Although easy to implement, such requirements perfectly. Both E2 and E3 match the
Santos et al.: HP Enterprise Services Uses Optimization for Resource Planning
158 Interfaces 43(2), pp. 152–169, © 2013 INFORMS
80
1.4
70
1.2
60
1.0
(%)
(%)
50
0.8
40
Satisfy
0.6
30 requirement 5%
0.4 of the time
20
0.2 10
0.0 0
1,800
3,600
5,400
7,200
9,000
10,800
12,600
14,400
00
00
00
00
00
00
0
30
80
30
8
,8
,3
,8
1,
3,
4,
6,
7,
9,
10
12
Figure 4: This graph shows that staffing projects based only on win probability may lead to poor service levels.
mandatory attribute skill group, but not all the other table with matching scores for less-than-100 percent
flexible attributes. For example, both E2 and E3 are matching. Both are used as inputs to the optimization
EXP for the attribute job level, which is one level module.
overqualified for the required INT level, and are
assigned with a similarity score of 0.25. Without The MIP Model
rescaling of weights (see Table 1), E2 and E3 will have We can verbally describe the MIP model in the opti-
the same matching score of 70, which cannot differ- mization module as follows. Appendix C provides
entiate between E2 and E3 for the job. After rescaling details about its notation and formulation.
(see Table 2), E2 is preferable to E3 because E2 has Decision variables: the MIP model makes the fol-
a higher rescaled matching score (62) than E3 (57.5). lowing decisions:
Note that E3 does not have the attribute of WF type, • employee allocation: a binary variable indicating
which is a more important attribute in the hierarchy if an employee is allocated to a job in a project oppor-
than other flexible attributes. tunity in a period;
Flexible matching output includes a qualification • hiring: number of people to hire for a job in a
table with 100 percent matching, and a transition project opportunity in a period;
Nonrescaled
Similarity score matching score
Resource attributes 45 4w · 5
Table 1: This table shows that without rescaling, E2 and E3 have the same matching score.
Santos et al.: HP Enterprise Services Uses Optimization for Resource Planning
Interfaces 43(2), pp. 152–169, © 2013 INFORMS 159
Nonrescaled
matching score Scaling coefficient 45 Scaled matching score
4w · 5 % 4 · w · 5
Downloaded from informs.org by [203.200.225.151] on 18 September 2016, at 06:05 . For personal use only, all rights reserved.
• training: a binary variable indicating if an — employee can be allocated to at most one job
employee allocated to a job in a project opportunity over the planning horizon
starts training in a period. — employee can only be trained once over plan-
Objective function: the objective function explicitly ning horizon
expresses a resource planner’s preference for allocat- — employee to be allocated (trained) must be
ing resources to fulfill demands. At HPES, job require- qualified and available
ments are first satisfied by available employees, who • other constraints
include both the current bench (i.e., employees cur- — tracking inventories of newly hired people
rently available for allocation) and future bench — hiring limits defined at different levels of
(i.e., employees who are currently assigned to ongo- aggregation.
ing projects but will be released at some period dur- Note that by allowing a gap for the unsatisfied
ing the planning horizon). If more than one employee demand, the MIP model will always be feasible. This
is qualified for a job, then the employee with the is a desirable feature for practitioners to use the RP
highest matching score is allocated. If no employee tool and interpret solutions.
is available or qualified to perform the job, then the
transition table is checked for an employee who can A Heuristic Solution Procedure
be trained to qualify for the job. If no employee can The MIP model is solvable using exact integer pro-
be trained, we consider hiring new employees. If hir- gramming methods such as branch and bound and
ing is impossible, we declare a gap. To summarize, branch and cut, which are available in many com-
the allocation preference is: current bench and future mercial solvers. Because the problem may involve
bench transition (training) joinee pipeline of new thousands of individuals and jobs, the size of the
hires gap. We model this allocation preference by MIP formulation can easily explode to millions of
appropriately specifying the job-resource allocation, binary decision variables and constraints. To make
training, and hiring penalties (costs) in the objective our optimization approach computationally tractable,
function. we designed and implemented a heuristic method to
Constraints: the main constraints of the MIP model rapidly obtain a near-optimal solution.
include The main concept of our heuristic is to solve a
• satisfying requirements of high-priority opportu- single-opportunity RP problem iteratively in a pre-
nities using available employees defined order, which is determined according to a
— through employee allocation or employee weighted priority score calculated from the oppor-
training tunity win probability, opportunity revenue, oppor-
— through hiring or identifying a gap tunity start time, and opportunity user-specified
• employee capacity and capability constraints priority. We sort the funnel of project opportunities
Santos et al.: HP Enterprise Services Uses Optimization for Resource Planning
160 Interfaces 43(2), pp. 152–169, © 2013 INFORMS
Demand Resource
No. of No. of No. of fulfillment utilization
Scenario # projects jobs employees Hiring limit (%) (%)
Downloaded from informs.org by [203.200.225.151] on 18 September 2016, at 06:05 . For personal use only, all rights reserved.
Table 3: This table shows a gap of the heuristic solution from the optimal solution for three scenarios.
in descending order of their weighted priority score. to fill the demand. The heuristic performs so well
The heuristic approach updates and solves a single- for the three scenarios because the resource supply is
opportunity MIP until all project opportunities— ample and not all the demand was considered: only a
including the planned bench—have been staffed. dozen high-priority project opportunities are staffed.
The heuristic approach typically takes 40–60 sec- To examine the heuristic’s performance under
onds to consider a dozen project opportunities with various resource capacities, we generate scenarios,
hundreds of jobs. The planned bench always takes adapted from real project data, considering expected
most of the run time, because it addresses supply utilization rates of 50, 75, and 95 percent. For each
and demand uncertainty; hence, it requires staffing expected utilization rate, we generate five instances
all the jobs that have an unfilled demand. We com- with workforce sizes ranging from 500 to 5,000 (see
pare the results of the heuristic approach with the Table 4). When we expect about half of the employ-
optimal solution for three real scenarios, which are ees to be committed, the heuristic solution deviates
randomly selected to show typical characteristics of from the optimal solution by about 6 percent on aver-
the RP problems at HPES. The scenario numbers age (for both the fulfillment rate and utilization rate).
are directly taken from the system. Table 3 provides Because of its greedy nature, the heuristic procedure
information about the three scenarios and the quality might assign an employee with two skills (one spe-
of the heuristic RP solutions. The last two columns cialized skill and one common skill) using his (her)
report the percentage of gap of demand fulfillment common skill for a higher-priority job. However, there
and resource utilization, respectively, of the heuristic might be a lower-priority job requiring the specialized
solution relative to the optimal solution. skill that cannot be staffed because capacity is tight.
For the three scenarios, the heuristic solutions are Thus, the heuristic tends to have lower fulfillment
optimal or near optimal—with gaps of less than and utilization rates when the resource supply is not
one percent. For Scenario 330, the solution is opti- ample. There are various ways to improve the heuris-
mal with respect to demand fulfillment, but not opti- tic approach. For example, when the more “flexible”
mal for resource utilization—with a gap of 0.57 per- employees (i.e., those who have more skills) are lim-
cent: the heuristic solution hires people to fill demand, ited in number, to avoid a shortage of skills, allocating
whereas the optimal solution is able to train resources less-flexible employees first would be better.
Table 4: This table shows a gap of the heuristic solution from the optimal solution when resource capacity varies.
Santos et al.: HP Enterprise Services Uses Optimization for Resource Planning
Interfaces 43(2), pp. 152–169, © 2013 INFORMS 161
from the funnel, the system does not use a signif- tion is unstructured, we cannot directly incorporate
icant amount of supply and demand information. it into the model and optimize allocation directly in
Although the model can optimize resource planning response. We cannot solicit this information because
we do not know its structure. Therefore, we allow
based on the information provided to the system, the
direct human intervention in the allocation space.
actual allocation is suboptimal because tacit human
That is, the process allows managers to adjust allo-
knowledge and judgmental information are not fully
cations. We created two specific mechanisms: preal-
incorporated. In this section, we discuss the specifics
location and gap allocation. An individual human
of information not captured in the system, and how
resource is frequently not unique within the RP
additional human processes that interact with the
system because the person has the same skill clas-
optimization system help capture some of this infor-
sification and salary levels as some peers. Thus, the
mation in the final allocation.
optimization problem that RP solves commonly has
Capturing human information is not a new idea
multiple solutions.
in the literature. For example, Plott and Sunder
In the preallocation mechanism, human managers
(1982, 1988) show that experimental markets can be
are encouraged to input allocations to projects (often
designed to solicit and aggregate human informa-
only a subset of all projects being allocated) before
tion into forecasts. This concept of a prediction mar-
RP finds an allocation solution. If the RP system finds
ket has been applied to contexts like political events
multiple solutions, it uses these preallocations to iden-
in the Iowa Electronic Market (Forsythe et al. 1992)
tify the most desirable solution; it uses only preallo-
and to sales forecasting at the Hewlett-Packard Com- cations that are consistent with RP optimization.
pany (Chen and Plott 2002). Many variations on the This process captures additional supply-side infor-
actual mechanism exist. Examples are Chen et al. mation such as “the manager knows that programmer
(2003, 2004), which use a combination of scoring rules A is more suited to do project X, although his skill
and behavioral modeling and correction. However, classification indicates that he is equally suitable for
we are not aware of any work, with respect to solic- X and Y .”
iting human information, in a planning setting. Thus, The situation in which certain positions cannot be
we had little theoretical guidance for designing an filled is almost as common as nonunique solutions.
information-soliciting mechanism that integrates with This happens whenever more positions than resources
a planning process. are available. Given the uncertain nature of both sup-
Financial estimates of opportunities and our like- ply and demand, this happens with a probability of
lihood of winning them are quantitative information one over an infinite horizon. When demand is larger
that the current system can capture. However, it can- than supply (i.e., a gap), RP will optimize the allo-
not capture tacit information such as the strategic cation with respect to financial objectives and match-
value of a deal (e.g., a small project that might lead to ing criteria. However, it cannot factor in strategic,
future opportunities). Similarly, managers might have often important, considerations. For example, a small
tacit information about their direct reports. For exam- project in a large account might not generate much
ple, two individuals, one from Mexico and one from revenue, but could result in a much bigger future
Iceland, might be classified as proficient in Java and opportunity.
be at the same job level; however, they might have Therefore, we created a gap allocation mechanism
different competencies relevant to a project opportu- to incorporate missing information. If the system iden-
nity. If an opportunity is in Mexico, a resource man- tifies a gap between available resources (supply) and
ager might know that the Mexican employee is more positions (demand), it flags the appropriate manager
familiar with the Mexican culture. Although the sys- and lets that manager decide which position should
tem can capture employee nationality, it cannot cap- not be filled; thus, managers can judge the relative
ture knowledge about such cultural competency. importance of projects beyond their financial impacts.
Santos et al.: HP Enterprise Services Uses Optimization for Resource Planning
162 Interfaces 43(2), pp. 152–169, © 2013 INFORMS
planning process, the process captures both the neces- represents a connection between execution and plan-
sary information flowing into the RP tool and how the ning. It is characterized by the current status of the
behavioral mechanisms (preallocation and gap alloca- resources and the expected labor demand for future
tion) are integrated into the system. First, we discuss periods within the cycle. Then the RP tool optimizes
the roles of resource managers and delivery managers. the resource allocation, transition, and hiring plans to
Resource manager functions include the following. maximize utilization and matching scores (i.e., mini-
• Update demand and supply data into RP in each mizes gaps). The user can define the cycle length; we
planning cycle by (1) entering labor requirements for set it as one week. When HPES wins a project, the
new or updated opportunities, and (2) classifying or delivery manager executes the allocation and transi-
updating resources by skills and business domain. tion plans. The next planning cycle then starts.
• Preallocate by designating preferred allocations.
• Run the RP planning scenario weekly by (1) de-
fining planning parameters (e.g., hiring bounds, ser-
RP Implementation and Benefits
vice levels), (2) executing the planning scenario We implemented the MIP model and heuristic pro-
through RP runs, (3) reviewing global metrics (i.e., cedure in GAMS (GAMS 2008). For the RP tool, we
resource utilization and demand fulfillment), and developed a graphical user interface (GUI), which
(4) running RP for unplanned demand, if needed. embeds the SDC and optimization modules at the
Delivery manager functions include the following. core. Using the Web-based RP tool, users can access
• Review allocation plans. Approved allocations the tool based on their roles, which can include
will be frozen as actions; open allocations will be dis- read and (or) write capabilities for certain data types
solved at the next RP run, and potentially new allo- and the ability to run the mathematical optimization
cations will be determined. engine. We initially deployed the RP tool at HPES in
• Resolve gap. The manger does this through a Bangalore, India in 2009.
manual allocation, by hiring, or by leaving a posi- Figure 6 shows screenshots of the allocation plan
tion open. and transition plan. The allocation plan advises deliv-
If HPES wins an opportunity or an unplanned ery managers regarding which employee will be
demand must be satisfied, delivery managers staff allocated to which job required by each project oppor-
the opportunity or unplanned demand using RP- tunity in the planning scenario (see Figure 6(a)). It
approved allocations. also defines periods of the planning horizons during
Input
Uncertainty
Demand Demand Demand
and labor and labor supply Demand
and labor supply and labor
forecast supply forecast
supply
Action RP RP
Today – 3 Today + 3
Planning horizon
Today
Planning Execution Planning
Execution
Soft Assignment Soft
Output Assignment
allocation and and project allocation and
and project
fulfillment schedule fulfillment
schedule
plan metrics plan metrics
Figure 6: These screenshots show the RP solution reports of the employee allocation plan (a) and the employee
transition plan (b).
Santos et al.: HP Enterprise Services Uses Optimization for Resource Planning
164 Interfaces 43(2), pp. 152–169, © 2013 INFORMS
which the allocation is active. Delivery managers may Conclusions and Future Research
approve, modify and approve, or leave open each In this paper, we address the workforce optimization
allocation in the plan. Approved, or modified and problem in a business setting by using a combination
Downloaded from informs.org by [203.200.225.151] on 18 September 2016, at 06:05 . For personal use only, all rights reserved.
approved, allocations are considered frozen alloca- of optimization mathematics and process design. The
tions. The RP tool cannot modify these allocations; it underlying problem involves staffing projects in the
can only modify an allocation that is left open. Fig- IT service industry using the right people at the right
ure 6(b) shows the matching score of an allocation. time and at the right cost. Many firms in the con-
If the matching score of the allocation is less than sulting and IT industries encounter similar problems.
100 percent, it will show the type of transition or train- We developed our RP optimization model to opti-
ing required to make the resource 100 percent quali- mize matching and allocation of employees to jobs in
fied for the job. a set of incoming project opportunities. By treating a
In comparison to the manual decentralized spread- project opportunity at an aggregate level, RP differs
sheet approach that resource and delivery managers from many existing project-staffing tools that focus
used previously, our RP tool has the following advan- on addressing assignment and scheduling decisions at
tages. First, it makes RP decisions in a centralized the execution level. To increase our billable resource
fashion, thus improving the visibility of the work- utilization, we consider workforce transition such as
force pool and providing resource economies of scale. training, promotion, or redeployment as an alterna-
Second, the automated RP process enhances the effi- tive to hiring new employees. To obtain reliable and
ciency and accuracy of staffing decisions. The RP robust solutions, we explicitly consider the dynamics
tool makes proactive decisions by considering var- and uncertainty from both the demand and supply
ious workforce demand and supply uncertainties sides. On the demand side, the win probability and
prior to making a decision; the manual approach uncertain starting times of project opportunities shape
was reactive—it made many last-minute decisions the demand pattern over time; on the supply side, the
in response to unexpected resource demand or sup- uncertain attrition rate directly affects the availability
ply. Moreover, the RP tool’s built-in optimization of the workforce.
functionality enables resource and delivery man- Because of the problem’s size and the need for fre-
agers to align their resource allocation decisions with quent RP decisions, we devised a two-phase approach
the strategic and tactical goals and business rules to model the RP problem. In the first phase, we
of HPES. Without optimization, resource allocation obtain estimated resource demand by computing the
decisions were often made subjectively and sponta- convolution of random variables given managers’
neously. For example, the managers who requested estimates about win probability, starting time, and
first or spoke the loudest often received the resources, attrition rate. We build buffer capacity into such esti-
to the detriment of more profitable project opportuni- mated demand (planned bench) to hedge against both
ties. In summary, the RP tool provides more objective demand and supply uncertainty, and then enter the
and better resource allocation solutions in terms of planned bench into the second-phase MIP model to
HPES internal workforce utilization, staffing cost, and obtain the optimal resource allocation, employee tran-
profitability. sition, and hiring plans. Our approach also includes
Since its deployment in our Bangalore operation, the a flexible matching method to allow and quantify
RP tool has enabled resource utilization rates of 90–95 imperfect matching between resources and jobs.
percent, compared with utilization rates of 75–80 In addition, we developed two accompanying
percent before its implementation. The Bangalore mechanisms to incorporate tacit knowledge and judg-
workforce of 5,000 is only a small portion of the world- ment, which the previous acquisition process did not
wide HPES workforce of 100,000; therefore, we expect capture, into our process. Although this is less opti-
significantly more benefit when we deploy the tool mal than incorporating the missing knowledge into
worldwide. We are currently working with our service the optimization model, it is a reasonable solution,
business unit for the ES-wide global deployment of RP given that we do not know the structure of the miss-
tool with suitable enhancements. ing knowledge.
Santos et al.: HP Enterprise Services Uses Optimization for Resource Planning
Interfaces 43(2), pp. 152–169, © 2013 INFORMS 165
The process is built into Web-based software. where 4i5 represents the duration of opportunity i. The
This software tool embeds modules for preparing summation ranges over all periods for which, if i starts in
input data, constructing and solving the optimization that period, it would still be active in period t. Thus, the
Downloaded from informs.org by [203.200.225.151] on 18 September 2016, at 06:05 . For personal use only, all rights reserved.
We calculate the scaling coefficient as follows: KjH : upper bound on number of workers hired for job j.
( K H : upper bound on total number of workers hired. The
a−1 if perfect matching1 hiring limits KjH and K H are defined by a resource planner
a = (B2)
Downloaded from informs.org by [203.200.225.151] on 18 September 2016, at 06:05 . For personal use only, all rights reserved.
where a−1 refers to the scaling coefficient of a’s immedi- C.2. Decision Variables
ate preceding attribute in the hierarchy, and the denomi- The main decision variables of the MIP model are as
P
nator f wf is the sum of original weights of all flexible follows.
attributes. The rationale behind such rescaling is that if a xwjit ∈ 801 19: xwjit = 1 if employee w is allocated to job j
resource misses an important attribute (with a large weight) of opportunity i in period t.
of a job, its matching score for the job should be penalized xwjit ∈ 801 19: wjit = 1 if employee w starts transition
(reduced). (training) for job j of opportunity i in period t.
hjit ≥ 0: number of people to hire for job j of opportunity
Appendix C. MIP Formulation i in period t.
jit ≥ 0: number of hired people allocated for job j of
C.1. Sets and Parameters opportunity i in period t.
I: set of project opportunities. The following additional derived decision variables are
J : set of jobs. needed.
W : set of available workforce. uwji ∈ 801 19: an indicator, uwji = 1 if employee w is used
T: set of periods in the planning horizon. for job j of opportunity i.
djit : number of workers required to perform job j of yjit ≥ 0: number of people needed that available employ-
opportunity i in period t. Resource and delivery managers ees cannot satisfy.
enter these data as soon as HPES decides to pursue the jit ≥ 0: inventory of people hired for job j of opportunity
opportunity. i in period t.
pi : win probability of opportunity i. The sales team esti- jit ≥ 0: gap or unfilled demand for job j of opportunity
mates and updates this value through the SOAR process. i in period t.
i : a flag; i = 1 means that an active opportunity i can
only hire new employees to fill the requirements. It applies C.3. Objective Function
to a technology project that requires new hires (e.g., college
graduates with a specific skill). min cwU ·
XXX
uwji
ljH : lead time for hiring a worker to perform job j. w j i
We obtain this from historical data. t
Qwj : qualification table. Qwj = 1 if employee w is 100 per- +c I ·
XXX
4jit −jit 5+
XX X
Rw −
XX
wjit
cent qualified to perform job j, 0 otherwise. j i t w t =1 j i
QTwj : employee transition or training table. Each entry in T
wjit +cjH ·
XXXX XXX
the table is the matching score 0 < wj < 1 of employee w +cwj · hjit
w j i t j i t
with job j.
T
lwj : lead time to transition employee w to perform job j.
+c L ·
XXX
It can be computed from historical data or provided by the 6−4t −157pi jit 0 (C1)
j i t
best estimates of resource planners.
Rwt : employee release period. Rwt = 1 means that
The objective function (C1) minimizes the total workforce
employee w is released from ongoing projects in period t;
cost (TWC) for service delivery, which consists of five terms.
we derive it from the status of the ongoing projects assigned
The first term is total utilization cost of available employ-
to w.
ees; the second term calculates total cost of idled employees,
cwU : cost of allocating employee w for a job in an oppor-
which includes both the unallocated new hires and future
tunity. We compute it from Equation (C3) with the base
bench (i.e., people released from ongoing projects); the third
allocation cost c0U set to 10 in our implementation.
and fourth terms represent the total cost of transition and hir-
c I : average cost for an employee to be idle. We set it at
ing, respectively; the fifth term calculates the expected loss of
50 in our implementation.
T the unfilled job opportunity period given the win probability.
cwj : cost for training employee w to be qualified for job j.
It is important to stress again that the objective function
We compute it using Equation (C4a).
can be configured to reflect the firm’s resource allocation
cjH : hiring cost for job j; we compute it using Equa-
preference. In our implementation, we enforce that
tion (C4b).
c L : average revenue loss because of a nonstaffed project cL cH > cT > cI > cU 1 (C2)
opportunity. We assign it a value far larger than the other
cost items. to encode the business rules at HPES.
Santos et al.: HP Enterprise Services Uses Optimization for Resource Planning
Interfaces 43(2), pp. 152–169, © 2013 INFORMS 167
Consider the number of periods a4w5 for which resource Constraint (C9) keeps the inventory-balancing constraint for
W has been waiting on the bench, and amax as the maximum new hires. X
number of periods for available resources. We calculate the xwjit ≤ T · uwji ∀ w1 j1 i0 (C10)
Downloaded from informs.org by [203.200.225.151] on 18 September 2016, at 06:05 . For personal use only, all rights reserved.
trained for the job, incurs a training cost higher than any ∀ w1 j1 i1 t0 (C11)
allocation cost of qualified resources. This way of comput-
Constraint (C11) specifies the conditions under which an
ing training cost is consistent with the hierarchy of pref-
employee can be allocated: (1) the hire-only flag must be
erence, which states that a qualified resource should be
off (i.e., i = 0); (2) an employee is qualified and avail-
allocated to a job before any trained resource is allocated able (the product of Rw and Qwj identifies whether an
to that job. Consider a base training cost c0T such that employee is qualified and released from ongoing projects
c0T = max8w ∈ W 2 cwU 9 + 1; the cost of training a resource can before period t); (3) or an employee can be transitioned and
be defined as the product of the base training cost times the available for transition (i.e., released from ongoing projects
T
lead time lwj to train the resource w for job j. Thus, we cal- before the period when the transition starts).
culate the training cost for resource w and job j as follows:
X
t
X
T
cwj = c0T · lwj
T
0 (C4a) QTwj · wjit · Rw ≤ 1 ∀ w1 j1 i0 (C12)
t =1
Similarly, to ensure that the training cost is always less Constraint (C12) restricts an employee to only one tran-
than the hiring cost, we compute a base hiring cost c0H = sition for a job in an opportunity.
T
max8w ∈ W 1 j ∈ J 2 cwj 9 + 1. Then the hiring cost cjH for job j is
hjit ≤ KjH ∀ j1
XX
(C13)
i t
cjH = c0H · ljH 1 (C4b)
hjit ≤ K H 0
XXX
(C14)
where ljH is the hiring lead time for job j. j i t
Chen, K., C. R. Plott. 2002. Information aggregation mechanisms: Valls, V., A. Perez, S. Quintanilla. 2009. Skilled workforce schedul-
Concept, design, and implementation for a sales forecasting ing in service centers. Eur. J. Oper. Res. 193(3) 791–804.
problem. Working paper, California Institute of Technology,
Yu, G., J. Pachon, B. Thengvall, D. Chandler, A. Wilson. 2004.
Pasadena, CA.
Downloaded from informs.org by [203.200.225.151] on 18 September 2016, at 06:05 . For personal use only, all rights reserved.
Haitao Li is an assistant professor of logistics and opera- maintenance policy. She earned a PhD in management sci-
tions management in the College of Business Administration ence at the University of Texas at Dallas in 2006.
at University of Missouri–St. Louis. His research interests Ravindra Kumar is affiliated with the HP Enterprise
Downloaded from informs.org by [203.200.225.151] on 18 September 2016, at 06:05 . For personal use only, all rights reserved.