P. 2 • IFORS NEWS December 2019
OR Impact
Articles demonstrating direct benefits from implementing OR studies
Automated Airport Staff Scheduling at Swissport
International Ltd.
Section Editors: Sue Merchant <suemerchant@hotmail.com>; John Ranyard <jranyard@cix.co.uk>
Prof. Dr. Andreas Klinkert <andreas.
klinkert@zhaw.ch>, Head OR/OM Group,
Technical Project Lead, Zurich University
of Applied Sciences (ZHAW)
Dr. Peter Fusek <peter.fusek@zhaw.
ch>, Lead Mathematical Modelling,
ZHAW
Bruno Riesen, Vice President Business
Support, Business Project Lead, Swissport
Roman Berner, Optimization Specialist, Swissport
Other Swissport Project Team members
Rita Thalmann, Lead Application Project
Simona Segessenmann, Application Specialist
Project Overview
Swissport International Ltd. is the world’s leading
provider of airport ground services and air cargo
handling, with 66,000 employees and a consolidated
operating revenue of 2.99 billion euros in 2018.
Swissport provided ground services for some 282
million passengers in 2018, and handled roughly 4.8
million tons of air freight in 115 cargo warehouses
worldwide. At the end of June 2019, Swissport was
active at 310 airports in 49 countries, on behalf of some
850 aviation customers.
Airport ground handling involves a broad range of
tasks, including passenger services like check-in, gate
handling and transfer, and ramp services like baggage
management, aircraft handling, and aircraft servicing
and cleaning. Swissport employs at its main airports
up to 2500 people with hundreds of dierent work
skills and shift duties, and a multitude of contract
types. Monthly sta planning is highly complex and
expensive, and usually requires extensive manual work
by specially trained personnel.
Swissport decided more than 10 years ago to start to
try to optimize these sophisticated planning processes
with the help of an appropriate software solution.
Evaluation of the commercially available sta rostering
tools showed that no software was able to satisfactorily
solve the complex large-scale planning problems at
Swissport.
Extracted from:
IFORS News
Volume 13, Number 4,
December 2019, ISSN 2223-4374
P. 3 • IFORS NEWS December 2019
Articles demonstrating direct benefits from implementing OR studies
In 2007, Swissport launched a strategic R&D cooperation
with the Zurich University of Applied Sciences Institute
of Data Analysis and Process Design (IDP), with the aim
of developing innovative high-performance software for
automated sta rostering that is capable of solving Swissports
challenging planning problems. In addition, the tool needed
to be suciently general and exible to be usable in other
companies and industries with complex rostering problems.
This collaboration, called Auto-Roster, started as a
Swiss national research project and has continued as
a strategic long-term cooperation between ZHAW and
Swissport, with an aim to continually adapt, extend,
improve and deploy the developed optimization tool.
Employee Scheduling
This generally comprises three main phases: (1) task
generation and shift construction, (2) rostering, (3) real-
time dispatching. Phase 1 includes demand modeling,
task generation, shift design, and demand covering.
Phase 2 is the main rostering phase where daily shifts
are assigned to individual employees. It consists of
the days-o planning and shift assignment tasks, and
generates a detailed monthly sta schedule which is
communicated to the employees several days before
execution. Phase 3 deals with the real-time planning
and control tasks on the day of operational execution.
The Auto-Roster project mainly focuses on Phase 2 which
represents the most complex, expensive and sensitive
planning task at Swissport. The amount of work involved in
rostering at Swissport can be illustrated with reference to
the initial situation at Zurich Airport, the pilot site for the
project. At project start, the rostering department comprised
20 planners who worked almost exclusively on rostering the
over 2000 employees at Zurich Airport which corresponds to
around 400 working days per monthly plan.
Major innovations and challenges for the Auto-Roster software
came from a dynamic, demand-driven planning policy which
does not rely on repetitive shift patterns rolled out over a long-
term horizon, and from a rostering approach which attributes
high importance to individual employee preferences. Human
planners typically achieved a 95% wish fulllment rate, and it
was imperative that an automated solution also met this rate.
Methodology
In contrast to most commercial rostering tools, which are
generally based on stochastic search metaheuristics, Auto-
Roster relies signicantly on Mixed Integer Linear Programming
(MIP), combined with various other optimization techniques,
including decomposition and relaxation, pre- and post-
processing, and a variety of heuristic procedures.
Approaching large-scale real-world rostering problems by MIP
techniques is innovative and challenging, since computation
times are typically far beyond any acceptable limits. Developing
good” MIP formulations to reduce solver computation times
was one of the most important and challenging parts of
the Auto-Roster project, and required deep knowledge in
combinatorial optimization, polyhedral combinatorics, and
graph theory. The project was several times close to failing
due to intractable MIP models and could only be continued
thanks to mathematical breakthroughs leading to powerful
new MIP formulations
Other major challenges came from feasibility issues, since
most real-world planning instances at Swissport are infeasible
at rst. Finding and explaining infeasibilities is intrinsically
complex, and a variety of algorithmic approaches had to be
developed to master these issues.
The current Auto-Roster optimization engine (without
front- and back-end) comprises some 25,000 lines of
MIP code and 30,000 lines of Java code. The largest MIP
instances contain around 1 million integer variables and
0.5 million constraints, and can typically be solved within
20 - 70 hours with relative MIP optimality gap << 0.5%. The
relative MIP optimality gap corresponds to the percentage
deviation of the best objective function value found so far
from the best objective bound resulting from the Branch-
and-Cut MIP solver process. For an illustration, see Fig.1,
which shows the convergence of the decreasing objective
function curve (measured in penalty points associated to
the current solution) and the increasing objective lower
bound, together with the associated MIP gap.