=Paper=
{{Paper
|id=Vol-2405/16_paper
|storemode=property
|title=Swarm Intelligence Layer to Control Autonomous Agents (SWILT)
|pdfUrl=https://ceur-ws.org/Vol-2405/16_paper.pdf
|volume=Vol-2405
|authors=Elnaz Khatmi,Wilfried Elmenreich,Kristina Wogatai,Melanie Schranz,Martina Umlauft,Walter Laure,Andreas Wuttei
|dblpUrl=https://dblp.org/rec/conf/staf/KhatmiEWSULW19
}}
==Swarm Intelligence Layer to Control Autonomous Agents (SWILT)==
https://ceur-ws.org/Vol-2405/16_paper.pdf
Swarm Intelligence Layer to Control
Autonomous Agents (SWILT)
Elnaz Khatmi1[0000−0002−1016−3933] , Wilfried Elmenreich1[0000−0001−6401−2658] ,
Kristina Wogatai1[0000−0001−5246−0672] , Melanie Schranz2[0000−0002−0714−6569] ,
Martina Umlauft2 , Walter Laure3 , and Andreas Wuttei4
1
Alpen-Adria-Universität Klagenfurt, Austria
elnaz.khatmi@aau.com,wilfried.elmenreich@aau.at,kristina.wogatai@aau.at
2
Lakeside Labs GmbH (LLabs) Klagenfurt, Austria
schranz@lakeside-labs.at,umlauft@lakeside-labs.com
3
Infineon Technologies Austria AG (IFAT)
Walter.Laure@infineon.com
4
Novunex GmbH (NoX)
andreas.wuttei@novunex.com
Abstract. The project SWILT focuses on swarms of cyber-physical sys-
tem (CPS)s in industrial plants (e.g., formed of products, machines, or
equipment). CPSs find their application in many disciplines including
Internet of Things (IoT), smart mobility, smart grids, Industry 4.0 and
smart houses. Swarms of CPSs are even more complex, hard to control
and program. To handle the complexity of swarms of CPSs, natural sys-
tems can serve as inspiration. Only through their interactions, a collective
behaviour emerges to solve complex tasks. SWILT considers the use cases
of production scheduling in industrial plants and transportation in logis-
tics. Currently, linear optimization is a widely used approach but due to
the increasing complexity it is typically performed only on a subset of
the industrial plant. Thus, current methods are unable to cope with the
search space of scheduling problems in large industrial plants. Since the
problem sizes in these use cases are extremely large and pre-calculated
schedules or transportation tables are not sufficient, the innovation is to
use swarm algorithms with reactive local rules on individual agents which
are able to compensate for dynamic system changes via local interactions
within their vicinity.
Keywords: Cyber-physical system · Swarm Intelligence · Self Organi-
zation.
1 Project data
Acronym: SWILT
Title: Swarm Intelligence Layer to Control Autonomous Agents
Start date: 1 October 2018
Duration: 36 month
Partners: Lakeside Labs GmbH (LLabs), Alpen-Adria-Universität Klagen-
furt (AAU), Infineon Technologies Austria AG (IFAT), Novunex
GmbH (Nox)
Website: https://swilt.aau.at
Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
92 E. Khatmi et al.
2 Introduction
Cyber-physical systems (CPSs) have strongly intertwined hardware and software
components, and find their application in many disciplines including IoT, smart
mobility, smart grids, Industry 4.0, and smart homes. In many cases, CPSs are
connected to other CPSs forming a system of systems or a swarm. Such swarms
of CPSs are even more complex, hard to control and program. This is reflected in
the current situation of the manufacturing processes for wafers. The production
of wafers is a highly dynamic process. Multiple machines need to be scheduled,
and repeated (between 400 and 1,200 different stations during a waferfab). This
results in an NP-hard problem when optimizing the WIP (work in progress)
flow, as there are nearly 2,000 different products. Another main challenge is to
integrate human work into optimized logistic processes.
SWILT focuses on swarms of CPSs in industrial plants (e.g., formed of prod-
ucts, machines, or equipment). To handle the complexity of swarms of CPSs,
natural systems can serve as inspiration. Therein, many homogeneous and het-
erogeneous agents cooperate without central control, executing simple rules lo-
cally. Only through their interactions, a collective behaviour to solve complex
tasks emerges. The SWILT concept embeds the local swarm rules in a three-
layered architecture: L3 - autonomous agents, L2 - swarm control, where each
swarm consists of a set of agents and their computation is intended to run as
5G network application, and L1 - central management (see Fig. 1).
3 Related Work
Other research activities combine the Particle Swarm Optimization (PSO) ap-
proach with other heuristics to improve its performance[15, 9]. The majority
of related work on the application of swarm concepts in production schedul-
ing build upon PSO or other swarm-based optimization algorithms. A notable
exception is the work by Leitão and Barbosa [8] using swarm agents to create
a self-organizing system for production scheduling. Another important contri-
bution to controlling swarms of CPSs comes from research on swarm robotics.
Fig. 1. The concept of applying a 5G network application.
Swarm Intelligence Layer to Control Autonomous Agents (SWILT) 93
Here, many simple small robots are coordinated in a self-organizing way follow-
ing the properties of a swarm. While approaches to define an intended behaviour
of a robot swarm [4, 13, 5] are inspiring for the research in this project, practical
implementations of robot-swarms are mostly on an experimental basis [1, 10] or
with an educational focus [7]. Overall, the state of the art shows that direct
applications of swarm intelligence (other than using swarm intelligence in an
optimization process), are very uncommon.
4 SWILT Project
SWILT aims at a direct application of swarm intelligence in industrial envi-
ronments. SWILT performs agent-based swarm modelling of an industrial plant
on the use cases of production scheduling and transportation in logistics. Since
the problem sizes in these use cases are extremely large and traditional pre-
calculated schedules or transportation tables are not sufficient, the innovation is
to use swarm algorithms with reactive local rules on individual agents which are
able to compensate dynamic changes in their local vicinity. The project will iden-
tify a library of suitable algorithms, define a model for intra- and inter-swarm
communication, and will show how to apply scenario-specific swarm intelligence
algorithms and how to extend swarm approaches with human-in-the-loop con-
cepts. To handle the complex communication within an industrial environment
with a huge number of agents and sensors, SWILT will elaborate the features
of 5G for handling the complex communication, such as direct device-to-device
communication and explicit support for communication intelligence at the edge.
The still premature communication technology promises to be an ideal match
for the SWILT layer concept and swarm communications in such an Industry
4.0 application.
4.1 Goals
To apply swarm intelligence algorithms to cope with NP-hard problems, the in-
dustrial plant must be modelled as a swarm of agents, where a set of components
from the same type, the same category or the same objective can be interpreted
as the swarm. A swarm consists of a large number of simple agents who together
pursue a specific goal based on local decisions of the agents. Since multiple com-
ponents interact, the goal is to construct multiple interacting swarms. Many
swarm intelligence algorithms have already been introduced in literature, but
they are rarely used as local algorithms in industrial plants. In a first step, a
theoretical analysis of these algorithms is performed in order to test whether
they are suitable for application to swarms in industrial plants. In this analysis,
the requirements of the application case and the requirements of the swarm al-
gorithm are included. Suitable algorithms are collected, the associated boundary
conditions and requirements documented and adapted to an extensible library of
swarm algorithms for use in industrial plants. Such a library allows us to reuse
algorithms, reproduce future results and achieve higher complexity goals. With
the help of the defined library for swarm algorithms, further evaluations can be
done. Another goal is to find a suitable simulation environment. Challenges here
94 E. Khatmi et al.
are to define a simulation that models the problem accurately enough so that
solutions that are elaborated based on information from such a simulation also
work in the real world, bridging the reality gap. On the one hand this calls for
a rather detailed and accurate simulation model, on the other hand potential
methodologies such as evolutionary methods involve a high number of simula-
tions which requires to complete thousands of simulations in short time. Poten-
tial algorithms are implemented, tested and analysed, to evaluate the effects of
swarm algorithms for both use cases. Achieving this aim gives us the ability to
make concrete statements on the usage of swarm algorithms in the industrial
domain. SWILT also envisages several swarms that communicate in different di-
rections, i.e., swarm2swarm, swarm2human and swarm2central communication.
A related goal is to define the best suited type or combination of communica-
tion technology. In particular, SWILT will take 5G into account and derive a
communication plan according to the specific characteristics of 5G. Only with
the application of 5G the SWILT swarm control layer L2 is able to run totally
independent from the underlying environment, to process the data on the edge
as network application and to handle the high amount of data traffic produced
through the mass of agents in the use cases.
4.2 Use Cases
The first use case addresses scheduling in semiconductor production systems.
This use case is of interest to the application of swarm algorithms, because cal-
culating optimal schedules is typically out of reach for such large-scale domains.
In wafer production, weekly workloads can involve around 105 operations on 103
machines. [12]
The main issues are to balance local constraints (for example, a machine
might process lots in batches and thus prefer to wait until a batch is (almost)
full) with competing constraints such as avoiding starvation of processing and
global objectives such as maximizing throughput. In complex processes such as
wafer production, a mixture of different products, dynamic changes in the system
and a high number of processing steps and involved machines form a scheduling
problem that due to its complexity can not be solved by exhaustive search during
production.
Another use case emerges from logistics in industrial plants, where the in-
tegration of human work together with automated systems forms a major chal-
lenge. Due to limited predictability and possibly limited compliance, an optimum
solution including the human factor is expected to be significantly different from
a logistic schedule of a completed automated system. Here swarm system are ex-
pected to provide the necessary flexibility and adaptability to integrate human
work successfully.
4.3 Methodology
The selection of algorithms and the modelling of the industrial plant are per-
formed upon the requirements/constraint analysis from the use cases. For an
initial test and analysis of these algorithms a common, easily programmable
Swarm Intelligence Layer to Control Autonomous Agents (SWILT) 95
simulation environment will be established that allows for fast evaluation of
algorithms and exploration of the nature of the problem. A simplified model
that still covers the main characteristics can also be used as a benchmark for
potential solutions where existing benchmark problems [11, 3, 14] are not spe-
cific enough. Moreover, an artificial test case also allows to be made public in
order to enable a reproducible evaluation of algorithms [6]. Possible platforms
for a simple simulation model are implementations in Netlogo5 , MATLAB6 , or
common programming languages with extensions for complex networks such
as Python7 /NetworkX8 . Based on the initial simulation model, possible candi-
dates for swarm algorithms will be evaluated. Besides the general paradigm (e.g.,
slime mold behavior, animal swarms, or other biological systems [2]) modeling
of swarm agents and the fine-tuning of algorithm parameters are issues that will
be addressed.
In parallel, a detailed concept for the data abstraction and communication
will be defined in form of a catalogue including the requirements and constraints
of the use cases. The concept will also take characteristics of communication
technologies into account, e.g., from the upcoming 5G standard, in order to select
the most appropriate technologies for intra- and inter-swarm communication.
Based on the use cases and requirements, test scenarios and experiments
are performed to test different swarm intelligence algorithms and the commu-
nication framework. In particular, tests and analysis are related to i) the ap-
plied algorithms, ii) their convergence to a defined goal (related to require-
ments/constraints), iii) the quality of inter- and intra-swarm communication
(data abstraction layers). The resulting algorithms will be evaluated quantita-
tively in a simulation based on the selected performance metrics. In addition, the
solutions will be reviewed by domain experts in order to assess their applicability
in a real productive setup.
5 Conclusion and Future Work
This paper introduced the SWILT project and laid out the basic design con-
cepts that are pursued by the project. The novelty of the SWILT project is the
application of swarm algorithms as a solution for coordination and scheduling
problems beyond the common application of swarm algorithms for optimization.
Thus the swarm members will be identified from hardware and software ele-
ments that are already available in the CPS. Within SWILT, example use cases
are production scheduling in semiconductor manufacturing and transportation
problems in industrial plants that take human operators into account. The main
contribution of SWILT, besides application in the specific use cases will be the
provision of a general architecture supporting multi-swarm systems with com-
munication models within and between swarms and to provide means for the
management of such swarm systems.
5
https://ccl.northwestern.edu/netlogo/
6
https://www.mathworks.com/products/matlab.html/
7
https://www.python.org/
8
https://networkx.github.io/
96 E. Khatmi et al.
Acknowledgments
This work was performed in the course of project SWILT (Swarm Intelligence
Layer to Control Autonomous Agents) supported by FFG – IKT der Zukunft
under contract number 867530.
References
1. Arvin, F., Murray, J., Zhang, C., Yue, S.: Colias: An autonomous micro robot for
swarm robotic applications. International Journal of Advanced Robotic Systems
11 (2014)
2. Camazine, S., Deneubourg, J.L., Franks, N.R., Sneyd, J., Theraulez, G., Bonabeau,
E.: Self-organization in biological systems. Princton University Press (2001)
3. Demirkol, E., Mehta, S., Uzsoy, R.: Benchmarks for shop scheduling problems.
European Journal of Operational Research 109(1), 137–141 (1998)
4. Dorigo, M., Trianni, V., Sahin, E., Groß, R., Labella, T.H., Baldassarre, G., Nolfi,
S., Deneubourg, J.L., Mondada, F., Floreano, D., Gambardella, L.M.: Evolving self-
organizing behaviours for a swarm-bot. autonomous robots. Autonomous Robots
p. 17:223 (2004)
5. Elmenreich, W., D’Souza, R., Bettstetter, C., de Meer, H.: A survey of models
and design methods for self-organizing networked systems. In: W. Elmenreich,
R. D’Souza, C.B., de Meer, H. (eds.) In Proceedings of the Fourth International
Workshop on Self-Organizing Systems. vol. LNCS 5918, p. 3749. Springer Verlag.
(2009)
6. Elmenreich, W., Moll, P., Theuermann, S., Lux, M.: Making computer science
results reproducible - a case study using gradle and docker. PeerJ Preprints 6
(2018)
7. Jdeed, M., Zhevzhyk, S., Steinkellner, F., Elmenreich, W.: Spiderino – A low-cost
robot for swarm research and educational purposes. In: 2017 13th Workshop on
intelligent solutions in embedded system. pp. 35–39. IEEE (2017)
8. Leitão, P., Barbosa, J.: Adaptive scheduling based on self-organized holonic swarm
of schedulers. IEEE 23rd International Symposium on Industrial Electronics (ISIE)
pp. 1706–1711 (2014)
9. Liu, B., Wang, L., Jin, Y.H., Tang, F., Huang, D.X.: Improved particle swarm
optimization combined with chaos. Chaos, Solitons & Fractals 25(5), 1261–1271
(2005)
10. Rubenstein, M., Ahler, C., Nagpal, R.: A low cost scalable robot system for collec-
tive behaviours. In Proceedings of the IEEE International Conference on Robotics
and Automation pp. 3293–3298 (2012)
11. Taillard, E.: Benchmarks for basic scheduling problems. European Journal of Op-
erational Research 64(2), 278–285 (1993)
12. Teppan, E.C.: Dispatching rules revisited – a large scale job shop scheduling ex-
periment. In: IEEE Symposium Series on Computational Intelligence (SSCI). pp.
561–568. IEEE (2018)
13. Trianni, V.: Evolutionary Swarm Robotics, vol. 108. Springer-Verlag Berlin Hei-
delberg, 1 edn. (2008)
14. Zhang, R., Wu, C.: A hybrid approach to large-scale job shop scheduling. Applied
Intelligence 32(1), 47–59 (2010)
15. Zhao, F., Zhu, A., Ren, Z., Yang, Y.: Integration of process planning and produc-
tion scheduling based on a hybrid PSO and SA algorithm. International Conference
on Mechatronics and Automation, Luoyang, Henan pp. 2290–2295 (2006)