On the Advancement of Project Management through a Flexible
Integration of Machine Learning and Operations Research Tools
Nikos Kanakaris, Nikos Karacapilidis and Alexis Lazanas
Industrial Management and Information Systems Lab, MEAD, University of Patras, 26504 Rio Patras, Greece
Keywords: Project Management, Machine Learning, Operations Research, Intelligent Optimization.
Abstract: Project Management is a complex practice that is associated with a series of challenges to organizations and
experts worldwide. Aiming to advance this practice, this paper proposes a hybrid approach that builds on
the synergy between contemporary Machine Learning and Operations Research tools. The proposed
approach integrates the predictive orientation of Machine Learning techniques with the prescriptive nature
of Operations Research algorithms. It can aid the planning, monitoring and execution of common PM tasks
such as resource allocation, task assignment, and task duration estimation. The applicability of our approach
is demonstrated through two realistic examples.
1 INTRODUCTION
Project Management (PM) is a complex practice that
is highly fluid and hard to predict, thus imposing a
series of challenges to organisations and experts.
Such challenges may concern alignment between
projects and their business objectives, handling of
conflicts and dependencies in resource allocation,
fine tuning of multiple projects to avoid fragmented
planning, as well as informed and diffused decision
making to handle potential opportunities or threats
during the execution of a project (Svejvig and
Andersen, 2015).
At the same time, PM is inherently collaborative
and knowledge-intensive. Issues to be addressed are
characterized by ever-increasing amounts of
different types of data and knowledge, which may be
obtained from various sources and vary in terms of
subjectivity, ranging from individual opinions and
estimations to broadly accepted practices and
indisputable measurements and results
(Karacapilidis, 2014). Their types can be of diverse
level as far as human understanding and machine
interpretation are concerned.
Up to now, the majority of methods and tools
aiming to facilitate and augment the quality of PM
are based on the application of advanced analytical
approaches developed and elaborated in the realm of
the Operations Research (OR) discipline. These
approaches employ techniques such as mathematical
optimization and statistical analysis to look for
optimal or suboptimal solutions to diverse PM
issues. In addition, the application of Artificial
Intelligence (AI) techniques to automate project
management has been proposed more than 30 years
ago. At that time, the proposed AI-leveraged project
management systems used knowledge processing
and procedural techniques to provide new kinds of
decision support for project objective-setting and
control (Levitt and Kunz, 1987).
Nowadays though, the adoption of AI in the
data-intensive and cognitively-complex PM settings
enables a series of advancements. AI, and in
particular Machine Learning techniques, can aid
project managers easily delegate thousands of tasks,
while sustaining a holistic view of their resources
and projects. This contributes to the achievement of
the required accuracy and precision when dealing
with bottlenecks or constraints that may obstruct
business processes. At the same time, these
techniques can aid managers and experts to interpret
big volumes of data and gain valuable insights
towards improving their overall PM practice. Based
on past data, they can predict undesired situations,
provide timely warnings and recommend preventive
actions regarding problematic resource loads or
deviations from business priority lists.
Admittedly, each of the abovementioned
disciplines (OR and AI) has significantly contributed
to the improvement of the PM practice, by
addressing the associated issues from a different
362
Kanakaris, N., Karacapilidis, N. and Lazanas, A.
On the Advancement of Project Management through a Flexible Integration of Machine Learning and Operations Research Tools.
DOI: 10.5220/0007387103620369
In Proceedings of the 8th International Conference on Operations Research and Enterprise Systems (ICORES 2019), pages 362-369
ISBN: 978-989-758-352-0
Copyright
c
 2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

philosophy and research perspective. However, we
argue that their joint consideration has not been
thoroughly explored yet, and has much potential to
further augment PM-related business intelligence.
Such an approach will concentrate on both the
planning and execution of individual projects as well
as their association with past data and their impact
on the wider business. Moreover, this approach can
appropriately represent and process the associated
data and knowledge, while at the same time remedy
the underlying cognitive overload issues. Particular
attention should be also given to the expression and
maintenance of tacit knowledge (i.e. knowledge that
employees do not know they possess or knowledge
that they cannot express with the means provided),
which predominantly exists and dynamically evolves
in PM settings.
In line with the above remarks, this paper
attempts to shape a hybrid approach for better
handling PM issues by meaningfully integrating
tools originally developed in the context of OR and
AI. The remainder of the paper is organized as
follows: Section 2 discusses background work
considered in the context of our approach, which is
analytically described in Section 3; the applicability
of the proposed approach is demonstrated through
two realistic examples presented in Section 4;
concluding remarks, limitations and future work
directions are summarized in Section 5.
2 BACKGROUND ISSUES
Numerous software solutions to PM exist in the
market nowadays. The list of the most widely
adopted ones includes Wrike (www.wrike.com),
Asana (www.asana.com), Trello (www.trello.com),
and Jira (www.atlassian.com/software/jira). Such
solutions offer a user-friendly environment that
mainly enables issue tracking and supports various
project management functions. In addition, by
providing interactive graphics, issue boards and
timelines, they simplify planning, collaboration,
reporting and time management. It is broadly
admitted that existing commercial PM solutions may
increase an organization’s productivity and prevent
the teams from diverging from their actual goals.
However, they unintentionally hide important PM-
related information, due to the complex
multidimensional data found in the hosted projects.
At the same time, by adopting an AI-perspective,
a range of digital project management assistants has
been already developed, including solutions such as
Stratejos.ai (www.stratejos.ai), PMOtto.ai
(www.pmotto.ai), and x.ai (www.x.ai). This
category of solutions is based on seamless, easy-to-
use interfaces that assist project managers in
common tasks (e.g. a project’s supervision). They
rely on the expressiveness, immediacy, interactivity
and descriptiveness that natural language provides to
offer a ‘zero-level’ entrance environment. They are
used to automate repetitive work such as creating
project’s tasks by analyzing textual conversations, to
remind and organize important events such as
meetings, to extract shallow insights (e.g. ‘top
contributors of the week’), and to answer simple
queries (e.g. ‘what is my team working on today?’).
We argue that this second category of solutions
offers narrow predictions and automations. In
particular, their underlying reasoning mechanisms
mainly build on rules to store and manipulate
knowledge, and ignore advanced AI technologies
that can uncover insights, perform more complex
tasks, make explainable recommendations, and
support informed decision making, sometimes in
ways that outperforms what people are able to do
today. Furthermore, each of these digital personal
assistants is relevant to a specific project
management need (e.g. reporting, scheduling
meetings, organizing events); thus, they are unable
to embrace a ‘single-access-point’ approach that
mitigates the overall PM complexity.
From an OR perspective, a series of techniques
and tools have been proposed and extensively used
to solve various PM related issues. OR techniques
provide solutions in problems such as prediction,
resource allocation, forecasting, scheduling, task
assignment, networking etc. These techniques are
supported by very useful software libraries such as
pyschedule (github.com/timnon/pyschedule), PuLP
(github.com/coin-or/pulp), Google OR-tools
(developers.google.com/optimization), JuMP.jl
(Dunning et al., 2017), Hungarian.jl
(github.com/Gnimuc/Hungarian.jl), and CVXPY
(www.cvxpy.org).
The abovementioned software libraries support a
variety of OR techniques including integer, linear,
convex and dynamic programming. However, these
techniques tend to add more complexity on the
overall PM practice, mainly due to the complicated
mathematical models needed to operate. Another
drawback is that these techniques are unable to learn
by the systems’ experience, which often results to
the proposition of optimal or near-optimal solutions
that are not realistically feasible.
With the advent of big data and cloud computing
era, Machine Learning (ML) techniques gain ground
in a variety of scientific and commercial sectors.
On the Advancement of Project Management through a Flexible Integration of Machine Learning and Operations Research Tools
363

These techniques (and corresponding algorithms)
can categorize items, predict values, identify
meaningful relationships, and detect data patterns or
unexpected behavior (anomaly detection). ML
approaches are usually grouped into four categories,
namely supervised learning, semi-supervised
learning, unsupervised learning and reinforcement
learning (Goodfellow et al., 2016).
Supervised learning refers to the process of
learning aiming to predict values (e.g. house prices)
or classify items into categories (e.g. categories of
projects) by using labelled training data. Common
algorithms and methods used in supervised learning
include k-nearest neighbors, naive Bayes, decision
trees, linear regression, and support vector machines.
Semi-supervised learning combines both labeled and
unlabeled input data for training, where in most
cases there is a small amount of labeled data and a
huge amount of unlabeled data available.
Unsupervised learning analyzes unlabeled data
to identify patterns or cluster similar items into
groups using alternative distance metrics (e.g.
Euclidean distance, Manhattan distance). Common
algorithms used in unsupervised learning include k-
means, DBSCAN, OPTICS, Apriori (Agrawal and
Srikant, 1994) and hierarchical clustering. Finally,
reinforcement learning approaches iteratively
interact with their environment to identify specific
actions that maximize the reward or minimize the
risk. Common algorithms and methods used in this
category include Q-learning, temporal difference,
and deep adversarial networks.
The above ML techniques and algorithms are
fully supported today by various software libraries
and environments, such as scikit-learn (Pedregosa et
al., 2011), H2O.ai (Candel et al., 2016), Tensorflow
(Abadi et al., 2016), PyTorch (Paszke et al., 2017)
and WEKA (Holmes et al., 1994).
As a last note, it is worth mentioning that most
AI-based approaches to PM build on artificial neural
networks. Related works discuss how neural
networks are capable to assist project managers in
problems such as resource allocation, prediction,
clustering, classification (Burke and Ignizio, 1992)
and forecasting (Zhang et al., 1998). Neural network
techniques have been also applied to predict
construction cost and schedule success (Wang et al.,
2012). Other representative works concern
development of a neural network to estimate project
performance (Cheung et al., 2006), or to classify the
level of a project's riskiness by exploiting the
knowledge extracted from data concerning past
successful and unsuccessful projects (Costantino et
al., 2015). An interesting overview of the different
types of neural network models applied in business
can be found in (Smith and Gupta, 2000).
3 THE PROPOSED APPROACH
Considering the pros and cons of the techniques
discussed in the previous section, we propose a
hybrid approach to handle contemporary PM issues,
which builds on a proper integration and
orchestration of PM tools originally developed
within the ML and OR disciplines. ML, which has
become a buzzword nowadays, adopts a predictive
analytics approach of the form ‘if A happens, then B
is likely to happen’, which attempts to exploit
available past data to create useful insights (i.e.
make human-like decisions). On the other hand, OR
adopts a prescriptive analytics approach to provide
optimal solutions (courses of action) to problems of
the form ‘what does A need to be if we want B to
happen’ (i.e. make perfect decisions).
We consider tools coming from the ML and OR
fields as complementary, arguing that there is room
for integration in a way that ML can create and
refine AB relationships that are often considered
as optimal and remain unchanged upon the entry of
new data in classical OR approaches. Despite the
features that ML possesses in terms of data
refinement and value prediction, it lacks algorithms
aiming to provide optimal solutions, something that
is inherent in OR techniques. Overall, our approach
considers OR and ML as complementary to each
other, and proposes an iterative interplay between
them, where ML supplies OR algorithms with
refined, accurate and up-to-date data (based on past
records), while OR contributes to making optimal
decisions with the continuously updated data input.
The proposed approach enables interpretation of
big volumes of PM data to support preventive
actions such as giving advice about resource
assignments by identifying similar skills and
expertise necessary to perform a task, make
explainable recommendations about the capacity
levels of certain resources based on historical
performance data, and support informed decisions
concerning a company’s expansion to a new region
or design of an efficient supply chain. The proposed
approach augments the overall PM decision-making
process, by enabling the drawing of reliable
conclusions about conditions and future events,
while also identifying potential risks and
opportunities.
Depending on the specific PM issue under
consideration, our approach advocates a proper
ICORES 2019 - 8th International Conference on Operations Research and Enterprise Systems
364

streamlining of ML and OR algorithms. As far as
ML algorithms are concerned, these can be
distinguished in four categories concerning data
classification, value prediction, structure discovery,
and detection of anomalies or abnormal behavior.
More specifically:
 Data classification aims to predict which
category the input data belongs to. For example,
in a software development project, a new task
can be classified into distinct categories (e.g.
story, bug, epic) based on its attributes using a
decision tree classifier.
 Value prediction concerns regression algorithms
to predict continuous numerical values. For
example, in a common PM scenario, these
algorithms can estimate the budget of a project
by exploiting knowledge of similar, already
accomplished projects using simple linear
regression techniques, thus providing advice to
the project manager during the planning phase
on possible cost reduction decisions.
 Anomaly detection algorithms aim to identify
unusual events or patterns that do not conform
to usual or expected behavior. For example, in a
certain maintenance setting, these algorithms
can detect outages of some components before
they occur and proactively act towards keeping
the whole system functioning.
 Structure discovery aims to uncover data
patterns, reveal hidden or not obvious
relationships and divide data items into groups
with similar traits (features). This is achieved
using widely-adopted ML techniques (e.g. k-
means and Apriori algorithms). For example, in
a construction PM problem, the Apriori
algorithm can mine frequent itemsets
concerning constructors and project durations to
build useful association rules (e.g. constructor x
is always late when delivering dam construction
projects).
4 EXAMPLES
In this section, we demonstrate the applicability of
the proposed approach through two realistic
examples concerning resource assignment. Emphasis
is given to the complementarity of ML and OR
algorithms to advance the associated PM practice.
Example 1
Based on real data concerning implementation of
public construction projects in the Region of Attica,
Greece, for the period 2003-2014, we consider the
following problem: Let P = {P
1
, P
2
, …, P
n
} be a set
of future public construction projects. Each project
(P
n
) is described by a list of attributes, namely P
n
=
[PID, location, category, est_cost, funding_source,
duration], corresponding to a unique project
identifier, the municipality to manage the project,
the type of construction needed, the project’s
estimated cost, the source funding the project, and
its estimated duration, respectively.
Similarly, let C = {C
1
, C
2
, …, C
m
} be the set of
registered construction companies, each of them
being associated with the set of attributes [CID,
{Location
i
}, {Category
j
}, {Cost_Range
k
},
{Duration_Range
l
}, AvgDiscount, AvgDelay],
corresponding to a unique constructor identifier, the
municipality where the constructor is active, the type
of projects the constructor deals with (e.g. flood
control, health infrastructure), the projects’ budget
category the constructor is interested in (e.g. large
scale (>1,5M€), medium scale (0,5M€-1,5M€)), the
projects’ duration range (e.g. short term (<6
months), mid term (6-18 months)), the average
discount provided by the constructor, and the
average delay caused by the constructor,
respectively.
Let a project management scenario where there
are n = 3 projects of various categories and m = 3
available constructors. Obviously, each P
n
requires a
different expertise, while each C
m
possesses a
distinct number of skills based on their profile,
which is populated with attributes extracted from
past data. To determine the constructors that best fit
to the projects’ requirements, we need to populate a
(P
n
, C
m
) score matrix (each entry taking values in the
range [0, 1]). This is through the calculation of (i)
the Jaccard similarity index J(P
n
, C
m
) (Jaccard,
1901), and (ii) an additional score value Score
C,M
for
the attributes avg_discount and avg_delay of each
C
m
(these attributes do not participate in the
calculation of the Jaccard similarity index).
We define:
Rating
n,m
= [ J(P
n
,C
m
) + Score(C
m
)] / 2
(1)
J(P
n
, C
m
) = | P
n
∩ C
m
| / | P
n
 C
m
|
(2)
Score(C
m
) =
[discount_score(AvgDiscount_C
m
) +
delay_score(AvgDelay_C
m
)] / 2
(3)
On the Advancement of Project Management through a Flexible Integration of Machine Learning and Operations Research Tools
365

discount_score




















(4)
delay_score =











 


 





(5)
By using formulas 1-5, we calculate the (P
n
, C
m
)
score matrix (Table 1).
Table 1: The (P
n
, C
m
) score matrix (Rating
n,m
).
P
1
P
2
P
3
C
1
0.8
0.8
0.3
C
2
0.6
0.7
0.5
C
3
0.7
0.4
0.8
Aiming to minimize the total construction cost of
these projects, the problem is considered as a typical
linear assignment problem (LAP), which can be
easily solved through tools available in widely used
software packages such as Google OR-Tools
(https://developers.google.com/optimization/assign
ment/simple_assignment). Using the linear
assignment solver of the above software package,
we get the outcome presented in Table 2.
Table 2: (P
n
, C
m
) assignment matrix.
Project
P
1
P
2
P
3
Constructor
C
1
C
2
C
3
Aiming to further improve the accuracy of our
estimations, we next consider the exploitation of
Machine Learning algorithms, which are capable to
provide knowledge-based patterns of construction
projects’ data. More specifically, we propose the use
of the Apriori Algorithm to discover meaningful
patterns (itemsets) relating P
n
and C
m
attributes.
Due to the fact that the Apriori algorithm
requires a full search of the transactions’ database in
order to generate a k-large itemset, we limit our
search to transactions containing only constructor
C
1
. We consider the transaction set T = {T
1
, T
2
, …,
T
16
} from a total of 685 transactions available in our
database, concerning constructor C
1
. The outcome of
Apriori algorithm provides us with a “strong”
supported 4-itemset that has been generated for
constructor C
1
(see Table 3; it is noted that, due to
space limitations, we present only the final step of
the algorithm, omitting intermediate calculations of
k-itemsets).
Table 3: L
4
itemsets for constructor C
1
.
Large Itemset (L
4
)
Support
BUILDINGS, LARGE_SCALE,
MID_TERM, DELAY_LEVEL_0
3
In order to construct the association rules for C
1
, we
define a set of rules R = {{R
1
, Conf(R
1
)}, .. {R
i
,
Conf(R
i
)}}, where:
R
i
= {X}  {Y}, where {X}, {Y}  {L
4
}
(6)
Conf(R
i
) = {X  Y}/{X}
(7)
According to Equations (6) and (7), the set of
rules produced is: R = {(R
1
, 1), (R
2
, 1), (R
3
,1), (R
4
,
1), (R
5
,1), (R
6
, 0.75), (R
7
, 0.75), (R
8
, 0.75), (R
9
, 1),
(R
10
,1), (R
11
,1), (R
12
, 1), (R
13
, 0.75), (R
14
, 0.75)}
For each project P
n
we apply the R
i
, where {P
n
∩
R
i
}≠, and we calculate the corresponding
confidence of the rule’s application.
We define:
Conf(RiPn ) =
(P
n
∩ R
i
/ R
i
) ∙ Conf(R
i
)
(8)
All R
i
with Conf(R
i
)≥0.5 are considered as
legitimate association rules and can be applied to the
initial available construction projects. In our case,
the rule R
9
with Conf(R
9
) = 0.5 has been applied to
P
2
project (Table 4).
Table 4: Applying R
i
 P
n
.
P
n
R
i
{P
n
} ∩ {R
i
}
Confi-
dence
(R
i
)
Prediction
(R
i
P
n
)
P
2
R
9
{BUILDINGS,
LARGE_SCALE}
0.50
{MID_TERM,
DELAY_LEVEL_0}
We notice that in P
2
project (originally assigned
to constructor C
2
– Table 2), the application of R
9
rule suggests (with high confidence) that an
assignment augmentation should take place. In other
words, Prediction(R
9
 P
2
) denotes that: “If C
1
constructor is selected for LARGE_SCALE
BUILDINGS, there is a 50% possibility to complete
P
2
project in MID_TERM duration and
DELAY_LEVEL_0 delay time”.
Taking into consideration the above prediction,
we update the original assignment matrix (Table 2).
The new assignments are shown in Table 5. We note
that for construction project P
2
={1002,
ATHENS,BUILDINGS, 3,67M€, EU, 1050 DAYS},
the above augmentation has a positive estimated
ICORES 2019 - 8th International Conference on Operations Research and Enterprise Systems
366

impact as it: (i) reduces its overall cost by 34.2%
based on C
1
profile (Avg_Discount
C1
), (ii) provides
minimum construction delay (DELAY_LEVEL_0) in
the range [0, 0.05] and (iii) reduces the construction
duration to MID_TERM (duration < 700 days).
Table 5: Augmented (P
n
,C
m
) assignments.
Project
P
1
P
2
P
3
Constructor
C
2
C
1
C
3
Our approach is sketched in a pseudo-code form
bellow:
for each (C
m
)in transactions_DB do
create_profiles(C
m
);
for each (P
n
) in projects_DB do
{
find top-n(C
m
);
calculate_Score(P
n
,C
m
) matrix;
}
assign(P
n
,C
m
);
for each C
1
in transactions_DB do // Apriori
// Algorithm
{
generate large k-itemsets (L
k
) with
minimum support (s);
construct rules (Ri) with minimum
confidence(Ri);
}
for each P
n
in projects_DB do
{
for each R
i
do
calculate_prediction(R
i
,P
n
);
}
assign(P
n
,C
m
);
To summarize the basic concepts of the above
example, we addressed a PM issue as a typical OR
assignment problem (a group of constructing
companies need to accomplish a set of construction
projects) using a score matrix with estimations for
each (Pn, Cm) element. LAP solver algorithm
provided a solution for the problem prescribing the
optimal assignment matrix. Next, we exploited ML –
Apriori algorithm to discover association rules
between transactions’ data to spot trends,
relationships and structure similarity between data
sets. In this way, we demonstrated that ML models
and algorithms can be used to re-feed initial OR
solutions, integrating OR’s prescriptive analytics
with ML’s predictive analytics orientation.
Example 2
Consider another project management scenario
concerning a software house, where there is a set
E={e
1
, e
2
, …, e
n
} of available employees (i.e.
software engineers). Each employee is described by
an ID and an array of skills. In addition, there are
two sets of new and past (completed) short-term
tasks (e.g. fixing of software bugs), that are denoted
by N={n
1
, n
2
, …, n
n
} and P={p
1
, p
2
, …, p
n
},
respectively, which are described by an array of
attributes, namely [description, assignee, skills,
duration].
By exploiting existing knowledge arising from
past similar tasks, the company desires to minimize
the total amount of time required to complete the N
new tasks. This process can be accomplished
through the following steps:
 Finding top-K employees for each task;
 Discovering clusters of similar tasks;
 Estimating tasks’ durations, considering
candidate employees;
 Assigning employees to tasks, by adopting the
linear assignment problem (LAP) algorithm.
More specifically, the company, for each new
task, discovers the top-K suitable employees by
comparing tasks’ requirements with each
employee’s skills. Obviously, for a specific task, the
most capable employee is the one who meets all the
required skills. Given a set X of a task’s required
skills and a set Y of an employee’s identified skills,
we define a score function  



as:






  




(9)
Next, our approach uses the k-means clustering
method to compose groups of similar tasks. The
features (attributes) used in our case include
[description, assignee_skills, task_skills]. It is
known that data pre-processing and preparation are
two fundamental steps in order for the k-means
algorithm to work properly. Hence, certain data
preparation techniques are applied to the features of
each task; these include (i) calculation of tf-idf
weights, removal of stop-words and stemming
regarding textual data (e.g. the description feature),
and (ii) conversion of an array of skills to binary
values (e.g. the skills features). It is noted that our
approach adopts the Elbow method (Trupti and
Prashant, 2013) to determine the number of k
groups.
For the abovementioned clustering requirements,
we use the scikit-learn package (which is
characterized by a wide adoption, simplicity,
usability, well-written documentation and code
stability). From the outcome of this step, we can
estimate the time t
ij
that each employee i needs to
complete a task j, which is equal to the time of the
task’s group centroid. The output of this step is shown
On the Advancement of Project Management through a Flexible Integration of Machine Learning and Operations Research Tools
367

in Table 6.
Table 6: Estimation of task durations per employee
(minutes).
N
1
N
2
N
3
E
1
118
‘N/A’
‘N/A’
E
2
63
546
116
E
3
287
179
184
E
4
‘N/A’
245
587
Applying the LAP algorithm
(https://developers.google.com/optimization/assign
ment/simple_assignment) to the elements of Table 6,
the final step assigns employees to tasks. The
outcome of this step is shown in Table 7).
Table 7: The task assignment matrix.
Task
N
1
N
2
N
3
Employee
E
1
E
3
E
2
It has been broadly admitted that employees
have the tendency to underestimate or overestimate
their skills, as well as the complexity of a certain
task in order to estimate the time or cost required to
accomplish it (Hill et al, 2000). As a consequence,
time estimations deviate significantly from the
reality, which in turn leads to miscalculations of
projects’ costs and durations, missing of deadlines,
etc.
By estimating tasks’ duration per employee
(Table 6) through the exploitation of past data, our
approach avoids the use of ad-hoc estimations and
feeds the LAP algorithm with more accurate input. It
is important to mention that the real duration of each
task is recorded (and compared to the estimated one)
for future use.
5 DISCUSSION
Key enablers that are driving the development of the
proposed approach are the availability of huge
computing power, the existence of big volumes of
PM data and knowledge, as well as the accessibility
of a range of well-tried and powerful OR and ML
software libraries. Undoubtfully, there is more
computing power available today than ever before,
something that contributes significantly in making
OR and ML algorithms extremely powerful, in ways
that were not possible even a few years ago. In fact,
this computing power enables us today to process
massive amounts of PM data and extract valuable
knowledge needed to make our models more
intelligent. At the same time, as discussed in Section
2, software needed to process the diversity of PM
data is open and freely available; it is also noted here
that PM-related AI algorithms become available and
get commoditized via dedicated APIs (Application
Programming Interfaces) and cloud platforms.
Despite the above advancements, much work
still must be done on the proper manipulation of PM
data and knowledge, as far as its labeling,
interrelation, modeling and assessment are
concerned; and this has mainly to be done by
humans. Especially in the context of project
management, one should always take into account
that valuable data and knowledge emerge
continuously during an organization’s lifecycle, and
concern both the organization per se (e.g. a project’s
duration, overall budget, KPIs etc.) and its
employees (e.g. one’s competences and
performance, knowledge shared during a decision-
making process etc.).
Building on a meaningful and flexible
integration of OR and ML techniques and associated
tools, our approach enables organizations to reap the
benefits of the AI revolution. It allows for new
working practices that may convert information
overload and cognitive complexity to a benefit of
knowledge discovery. This is achieved through
properly structured data that can be used as the basis
for more informed decisions. Simply put, our
approach improves the quality of PM practice, while
enabling users to be more productive and focus on
creative activities. However, diverse problems and
limitations still exist; these concern the value and
veracity of existing data, as well as the availability
of the massive amounts of data required to drive
contemporary AI approaches.
Future work will concentrate on the
consideration of more complex PM issues aiming to
identify additional useful combinations of ML and
OR algorithms. Another work direction concerns the
embedment of explainability features in the
recommendations provided by the proposed
approach (Karacapilidis et al., 2017).
6 CONCLUSIONS
Considering contemporary PM challenges as well as
the strengths of techniques originally developed in
the context of the ML and OR fields, this paper
presents a hybrid approach that aims to advance the
ICORES 2019 - 8th International Conference on Operations Research and Enterprise Systems
368

overall PM practice. The proposed approach can
assist employees in common PM tasks such as
resource assignment, estimation of task duration,
and prediction about whether deadlines will be met.
The proposed advancement of the PM practice lies
in the proper orchestration of OR and ML
algorithms by paying simultaneous attention to both
optimization and big data manipulation issues.
REFERENCES
Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A.,
Dean, J. and Kudlur, M. (2016). Tensorflow: a system
for large-scale machine learning. In Proceedings of the
12th USENIX Symposium on Operating Systems
Design and Implementation, 16(1), pp. 265-283.
Agrawal, R. and Srikant, R. (1994). Fast algorithms for
mining association rules, In: Proceedings of 20
th
int.
conference on Very Large Data Bases (VLDB 1215),
pp. 487-499.
Burke, L. I. and Ignizio, J. P. (1992). Neural networks and
operations research: An overview. Computer &
Operations Research, 19(3), pp.179-189.
Candel, A., Parmar, V., LeDell, E. and Arora, A. (2016).
Deep learning with H2O. 6
th
ed. [pdf] H2O.ai Inc.,
Available at http://h2o.ai/resources/ [Accessed 19 Oct.
2018].
Cheung, S. O., Wong, P. S. P., Fung, A. S. and Coffey, W.
(2006). Predicting project performance through neural
networks. International Journal of Project
Management, 24(3), pp. 207-215.
Costantino, F., Gravio, G. D. and Nonino, F. (2015).
Project selection in project portfolio management: An
artificial neural network model based on critical
success factors. International Journal of Project
Management, 33(8), pp. 1744-1754.
Dunning, I., Huchette, J. and Lubin, M. (2017). Jump: A
modeling language for mathematical optimization.
SIAM Review, 59(2), pp. 295-320.
Goodfellow I., Bengio Y. and Courville A. (2016). Deep
Learning, The MIT Press. Cambridge, MA, USA.
Hill J., Thomas L.C. and Allen D.B. (2000). Experts'
estimates of task durations in software development
projects. International Journal of Project
Management, 18(1), pp. 13-21.
Holmes, G., Donkin, A. and Witten, I. H. (1994). Weka: A
machine learning workbench. In Proceedings of the
1994 Second Australian and New Zealand Conference,
pp. 357-361.
Jaccard, P. (1901). Étude comparative de la distribuition
florale dans une portion des Alpes et des Jura. Bull Soc
Vandoise Sci Nat, 37, pp. 547-579.
Karacapilidis, N., ed. (2014). Mastering Data-Intensive
Collaboration and Decision Making: Cutting-edge
research and practical applications in the Dicode
project. Studies in Big Data Series, Vol. 5, Springer,
Cham.
Karacapilidis, N., Malefaki, S. and Charissiadis, A.
(2017). A novel framework for augmenting the quality
of explanations in recommender systems. Intelligent
Decision Technologies Journal, 11(2), pp. 187-197.
Levitt, R. E. and Kunz, J. C. (1987). Using artificial
intelligence techniques to support project
management. Artificial Intelligence for Engineering
Design, Analysis and Manufacturing, 1(1), pp. 3-24.
Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E.,
DeVito, Z. and Lerer, A. (2017). Automatic
differentiation in pytorch. In 31st Conference on
Neural Information Processing Systems, pp. 1-4.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,
Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P.,
Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
Cournapeau, D., Brucher, M., Perrot, M. and
Duchesnay, E. (2011). Scikit-learn: Machine Learning
in Python. Journal of Machine Learning Research,
12(1), pp. 2825-2830.
Smith, K. A. and Gupta, J. N. (2000). Neural networks in
business: techniques and applications for the
operations researcher. Computers & Operations
Research, 27(11), pp. 1023-1044.
Svejvig, P. and Andersen, P. (2015). Rethinking project
management: A structured literature review with a
critical look at the brave new world. International
Journal of Project Management, 33(2), pp. 278-290.
Trupti M. K. and Prashant R. M. (2013). Review on
determining number of Cluster in K-Means Clustering.
International Journal of Advance Research in
Computer Science and Management Studies, 1(6), pp.
90-95.
Wang, Y. R., Yu, C. Y. and Chan, H. H. (2012).
Predicting construction cost and schedule success
using artificial neural networks ensemble and support
vector machines classification models. International
Journal of Project Management, 30(4), pp. 470-478.
Zhang, G., Patuwo, B. E. and Hu, M. Y. (1998).
Forecasting with artificial neural networks: The state
of the art. International Journal of Forecasting, 14(1),
pp. 35-62.
On the Advancement of Project Management through a Flexible Integration of Machine Learning and Operations Research Tools
369