You are on page 1of 6

Available online at www.sciencedirect.

com

ScienceDirect
Procedia CIRP 119 (2023) 656–661

33rd CIRP Design Conference

A Living Lab Platform for Testing Additive Manufacturing Agent-Based


Manufacturing Strategies
Lorenzo Giuntaa,, Ben Hicksa , Chris Snidera , James Gopsilla,b
a School of Civil, Aerospace and Mechanical Engineering, University of Bristol, UK
b Centre for Modelling and Simulation, UK

* Corresponding author. E-mail address: l.giunta@bristol.ac.uk

Abstract
Distributed manufacturing systems are seeing increased investment by manufacturers looking for a more responsive, robust, and sustainable means
to meet product demand amid market fragility, mass-customisation, sustainability, supply chain uncertainty, national security, Net Zero and the
Circular Economy. One such system is Additive Manufacturing farms where numerous machines are housed in warehouses across the globe. The
realisation of these systems in industry is offering the opportunity to exploit research in agent-based manufacturing strategies for the brokering and
coordination of jobs. Until now, much of the underpinning research has been numerical and validation via empirical study is required to provide
confidence in deploying the strategies in an industrial setting. To provide this validation, this paper presents an Opensource Living Lab platform
that can be used to evaluate agent-based manufacturing strategies.
© 2023 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer review under the responsibility of the scientific committee of the 33rd CIRP Design Conference

Keywords: Distributed Manufacturing; Agent-Based Manufacturing; Living Lab; Validation

1. Introduction clude: Additive Manufacturing, robotics, automated inventory,


closed-box Computer Numerical Control (CNC), the Internet-
Manufacturing and Supply Chains are experiencing unprece- of-Things (IoT), networking, cloud infrastructures and man-
dented change [10]. Changes in societal behaviour, such as ufacturing resource planning software systems. Combined, it
the ‘Maker’ movement, mass-customisation, and the desire for enables a readily deployable flexible manufacturing capability
rapid delivery require an on-demand and responsive manufac- within standard warehouse infrastructure. Demand is then dis-
turing capability. Net Zero and the Circular Economy have de- tributed amongst the many machines housed in the system.
veloped an existential crisis in how society should manufacture Unlike conventional production, flexible distributed manu-
and supply itself [3]. Market fragility and supply chain uncer- facturing systems afford [1, 16, 7]:
tainty is increasing the risk in forecasts, resulting in smaller and
more intermittent orders being made. National Security agen-
das are encouraging the re-shoring of manufacturing and an in- 1. Reduction in distribution costs.
crease in global ‘shock’ events, such as COVID-19 and the war 2. Shorter lead times.
in Ukraine, are requiring more resilient manufacturing and sup- 3. Less capital investment.
ply chain architectures to be developed [18]. 4. More flexibility in capacity.
In light of these events, manufacturers have been increas- 5. Better capacity utilization
ing investment in the study and deployment of flexible dis- 6. Larger pool of experts.
tributed manufacturing systems brought about by a combination 7. Diffusion of risk.
of digital-physical manufacturing innovation. Innovations in- 8. Boost to local economies.
9. Sustainability.
✩ This document is the result of research funded by the United Kingdom’s En-

gineering and Physical Sciences Research (EPSRC) Council – EP/R032696/1, Additive Manufacturing (AM) is one of the manufacturing
EP/W024152/1 and EP/V05113X/1. processes driving Flexible Distributed Manufacturing Systems.
It affords the ability to rapidly change between component man-

2212-8271 © 2023 The Authors. Published by Elsevier B.V.


This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer review under the responsibility of the scientific committee of the 33rd CIRP Design Conference
10.1016/j.procir.2023.03.118
Lorenzo Giunta et al. / Procedia CIRP 119 (2023) 656–661 657

ufacture with little to no re-configuration and/or changeover 2. Requirements for a Living Lab to test Additive Manu-
time [6]. It can also be readily deployed in a range of loca- facturing Agent-Based Manufacturing System concepts.
tions from homes, schools, universities, offices and standard
warehouse infrastructure. For example, there are an estimated When considering the type of infrastructure to utilize for
168,000–230,000 3D printers in the United Kingdom [19, 20] the development of the LL, a number of architectures were
resulting in a combined manufacturing capability that could considered. PROSA, FABMAS, and ADACOR, architectures
produce 2.02–2.76 Million Lateral Flow Devices per day 1 . are based on the Multi Agent Systems (MAS) holons concept
The challenge for manufacturers wishing to unite and oper- [22, 14, 8, 4]. Holon, is derived from the Greek “holos” (whole)
ate a Flexible Distributed Manufacturing System is the broker- and the suffix “-on” to indicate a particle or subset [21]. As
ing and coordination of work. The systems must react to con- hinted at by the name holons are used to describe systems that,
tinually changing and unpredictable demand profiles that need due to their complexity, are built of increasingly small sub-
to be distributed amongst the numerous machines within the assemblies which remain fully functional microcosms. This du-
system. Machines may be in different locations and operated ality is how holons can be seen as wholes which exist due to
by different organisations, with differing requirements and con- smaller parts but also as interconnected and inter-operable as-
straints (e.g. availability of certain materials). This limits the ap- semblies. While not strictly hierarchical, holons tend to favour
plicability of typical top-down production control methodolo- hierarchical orders to ensure the interoperability of the system,
gies [9]. In these cases, Agent-Based Manufacturing strategies as smaller assemblies must fit within larger ones [4]. Thus the
have been identified as a suitable solution. Agent-Based Man- models (PROSA, FABMAS, and ADACOR) maintain a form
ufacturing represents machines and jobs as agents. The agents of hierarchy which is inherent to the holon philosophy.
feature the necessary information to represent a particular ele- As the network that the LL aims to replicate and test is one
ment of the system and operate autonomously communicating based on leveraging makers and disparate manufacturers, it is
with other agents. Through their communications, work is co- hard to create a hierarchical structure to ensure coordination.
ordinated and distributed. Agent-based systems increase flexi- This is particularly true for individual makers and outfits run-
bility and responsiveness due to the agents’ independence, al- ning just one or a few AM machines. In such a scenario, where
lowing for the parallelization of decisions and real-time adjust- there is low organisation across holons, MAS begin to more
ments [15, 5]. closely mimic the behaviour found in heterarchical systems,
However, many of the advantages have only been evidenced as the holon consists of a single AM machine [2]. Anarchic
through numerical modelling due, in part, to the lack of real- systems are a form of heterarchical production control systems
world capability to realise such systems. As is often the case [12], but their interaction is based on an open market approach.
with numerical modelling, they feature assumptions of ide- As with heterarchical systems, the control and governance is
alised conditions and do not necessarily concern themselves maintained at the agent-level. However, agents within anarchic
with the practical challenges of implementing their communi- systems are in a continuous state of competition as they bid for
cation strategy in a real-world setting. Example challenges in- work, mimicking the interactions found in stock markets, and
clude the consistency of operators in maintaining system oper- collaboration is obtained despite each agent working primarily
ations, robustness of the machines in operation, and the will- for its own self interest [11, 13].
ingness to share information between agents. It is argued that a While desirable to have such an approach for the final im-
empirical studies through real-world implementations is a con- plementation of the ABMS for AM, the added requirement for
tributing factor to the lack of adoption. the LL is to allow for data collection and an overview of the be-
To fill this gap, this paper details an Opensource Living haviour of the agents in the LL. Challenges in these systems is
Lab (LL) platform to support the field in testing and evaluat- how the local control and governance can be configured in such
ing Agent-Based Manufacturing strategies and thus help move a manner to provide the desired emergent system performance.
the technology along the Technology Readiness Levels (TRLs). Implementing and studying these systems requires fine-grained
The paper continues with a review of Agent Based Manufac- monitoring of the communications between agents in order to
turing System (ABMS) research where the requirements for a draw correlations between agent configurations and system be-
platform are elicited (Section 2). This is followed by a descrip- haviour.
tion of the platform’s architecture (Section 3). A demonstra- As such, the architecture selected for the LL is inspired by
tion of the platform with a published communication strategy both the ADACOR and anarchic architectures. From ADA-
is presented and comparison between real-world and simulated COR, the proposed LL architecture borrows both the use of
results made (Section 5). A discussion then details the insights separate holons connecting into a network infrastructure and
that can be generated through studies using a LL (Section 6). the use of a virtual resource to represent the physical manufac-
The paper then concludes with the key contributions of the work turing resources available (AM machines) which is necessary as
(Section 7). the machines in question are not designed to connect to such a
network and take jobs independently [8]. However, in line with
the flat hierarchies proposed by anarchic manufacturing [12] the
LL’s architecture eschews ADACOR’s hierarchical setup, pre-
ferring a flat organisation for both machine and job agents, with
1 2hrs printing time per device [17].
658 Lorenzo Giunta et al. / Procedia CIRP 119 (2023) 656–661

the network infrastructure acting merely as a passthrough con- the communication behaviour of agents and the emergent be-
necting the machines to the jobs. This central passthrough point haviour of the system. Authentication tokens and SSL encryp-
allows for an overview of the network to be maintained, which tion (via LetsEncrypt) are configured when the Broker service
would be complex if not impossible if all the communication starts providing security. The Broker also validates the mes-
between agents occurred in a distributed manner, as proposed sages being sent and will return an error if it does not conform
by fully anarchic systems. This allows for data collection and to the format specified. Agents connecting to the Broker also
network monitoring, favouring the LL’s ability to gain informa- need to specify the type of agent they are and the group they
tion on how various experimental parameters impact the output belong to.
and behaviour of the overall manufacturing system, allowing Listing 1. Message structure for communication between agents and broker
for comparison to computational models. For example, it would // A list of agent uuids that
be possible to explore how different protocols for job selection // the message will go to .
by the virtual resources impact the number of jobs manufac- to : string []
tured and the time each job waits before being accepted. // The agent uuid the message
// originated from .
from : string
3. A Living Lab for AM ABMS // A pre - defined subject enum .
subject : enum
Fig. 1 illustrates the template for the AM ABMS discussed // A key - value object conditional
in the previous section. This template formed the basis of the // on the subject .
message : {[ key : string ]: any }
design of the Living Lab’s architecture. The four components
highlighted therein were: Listing 1 shows the message structure used to communicate
with the broker. The ‘to’ field enables an agent to supply a list
1. Broker service (Network Infrastructure) of unique agent identifiers (e.g. uuidv4 format) that the mes-
2. Machine connectors (Physical Resource) sage should be sent to. An empty list indicates a broadcast mes-
3. Machine agent (Virtual Resource) sage, i.e. a message to be sent to all agents on the network.
4. Job representatives (Job Agent) The ‘from’ field contains the unique agent identifier of the orig-
inating agent. The ‘subject’ field provides a meta-description
of the type of message being sent (e.g. seeking jobs, accept-
ing jobs, etc.) in line with the schema provided on instantiating
the service. The ‘message’ field then contains additional neces-
sary information, such as gcode. This approach allows for the
integration of multiple types of agents, as they simply need to
follow the message structure.
The Machine connectors (2) provide AM machine connec-
tivity enabling the web app to report the current state of a ma-
chine as well as send gcode to the machine for manufacture.
Machine connectors include:

• Manual workflows where the connector provides access


to the job gcode for download for further processing.
• Direct serial connection to a machine.
Fig. 1. Living Lab Architecture. • Accessing of a machines Application Programming In-
terface (API) (e.g., Ultimaker API).
The entire stack is web-based (Typescript) so end-users do • Accessing a machine through a remote service API (e.g,
not need to install any software on their infrastructure, which is Octoprint API).
often a challenge with an organisation’s IT policies. • Accessing a Fleet Manager API (e.g. Eiger.io and Ulti-
The Broker (1) handles the communication between ma- maker Digital Factory) where the connector provides ac-
chine and job agents. The broker supports direct communica- cess to a set of machines.
tion between individual agents as well as broadcasting mes-
sages to all Job and/or Machine agents. The protocol used was The clients created thus far include an Ultimaker client that
socket.io, which uses the websocket protocol to establish a con- talks with the Ultimaker API, an OctoPrint client that talks with
nection between host and client. Socket.io permits multiple in- the Octoprint API, and WebUSB clients that provides a direct
stances of the Broker providing horizontal scaling to the ser- serial connection to a 3D printer (Links to all the required repos-
vice. The Broker also features a logging service that can log all itories may be found at: https://tinyurl.com/mhxp7dkm).
the messages sent between the agents. This is stored as a JSON The connectors enable a range of machines to be connected to
file on the server that can be accessed via a REST API. This a single network and provide the fundamental capabilities for
is a critical feature needed to study the impact of changes to sending machine code to be printed, reporting printer status and
Lorenzo Giunta et al. / Procedia CIRP 119 (2023) 656–661 659

cancelling prints if there is an issue, and setting a Boolean value met, the Job agents respond to the Machine agent via the Bro-
to indicate whether the machine is available, which in turn, per- ker. As a FRFS strategy was implemented, the machine will ac-
mits the machine agent to start looking for work. cept the first Job that messages it within the period of waiting.
Having connected the machines to the web app in the The Machine agent then responds with a "machine has chosen
browser, there needs to be a method for connecting the ma- job" message informing the Job agent that it was selected. The
chine to the service and act as its representative (agent) in the Job agent receives this message and checks whether, in the in-
system. The Machine agent (3) class handles the connection to tervening time, manufacturing was started by another Machine
the Broker, stores the metadata representation of the machine, agent. If the Job agent is still available, it will respond with a
selects the logic used to bid for work, and passes gcode to the "job has accepted offer" message that includes the gcode for the
Machine connector for printing. Different logics can be imple- Machine agent to begin manufacture.
mented though a set of functions that listen on message events
coming from the Broker.
The Job representative (agent) (4) handles the connection to
the Broker, stores the metadata representation of the Job, and
permits a set of gcodes to support printing on multiple types of
machine. The Machine connectors, Machine agent and Job are
combined into a web application that allows an end-user, inde-
pendent of hardware and OS, to initialise and add agents to the
network. Adding multiple agents to a system can be achieved
through multiple tabs on a single computer or across multiple
computers.
Fig. 2. Example communication response between jobs and a machine in the
Living Lab. Numbering indicates the order in which the messages are sent.
Starting at (1) the Machine agent begins the interaction by seeking available
4. Demonstrating the Living Lab Platform
jobs.

To demonstrate the functionality of the proposed Living Lab


In order to simulate a demand scenario to evaluate the effec-
architecture. The Living Lab was used to create an ABMS with
tiveness of the Living Lab, a script was created to periodically
four 3D printers: 2 Ultimaker 3 Extended and 2 Ultimaker S3 as
create Job agents. These would exist until a machine selected
a form of ‘shakedown’ test. These were connected to the server
the job for manufacture. For the purpose of this scoping study,
and assigned a coordinated manufacturing protocol. Five logics
the item to be manufactured was a cylinder. The gcode for this
were added to the system:
was generated in order to take approximately 10 minutes to
manufacture on both types of machines used. The service oper-
1. First Response First Serve (FRFS) – The Machine agent ated for a period of 3h:30m. All communications made via the
accepts the first Job agent that responds to a request to be Broker were logged.
manufactured. The log was post-processed to obtain the demand profile
2. Shortest Manufacture Time (SMT) – The Machine agent generated by the job submission script over the 3h:30m dura-
accepts the Job agent with the shortest manufacture time. tion of the study. This demand profile was then replicated and
3. Random Allocation (RA) – The Machine agent accepts a inputted into an agent-based numerical model. The numerical
Job at random. model was then run using the same demand profile.
4. Longest Manufacture Time (LMT) – The Machine agent The results of the model were then compared to the print
accepts the Job agent with the longest manufacture time. start times extracted from the logs and a comparison was made
5. First Come First Serve (FCFS) – The Machine agent ac- between the two. The differences between expected print start
cepts the Job agent that first entered the job pool. times, as predicted by the model, and actual start times of each
print highlights the inaccuracies and/or simplification of the
Of the five considered protocols for job selection FRFS was model. This allows for a better understanding of the “dead time”
selected. Fig. 2 provides an example of a typical communica- that cannot be directly accounted for, such as the time required
tion between the Machine and Job agents through the Broker. to: clear the AM machine of the previous print, cleaning, cool-
While the Machine agent is in the available state, the Machine ing/heating the bed/nozzle(s), etc.
agent will periodically send a message informing the Broker
that it is looking for jobs. The Broker will relay this message
to all Job agents connected to the network. The Machine agents 5. Results
will then enter a period of ‘waiting for responses’ from the job
agents. One observation from the setup of the Living Lab was the
The Job agents will respond only if they have not been man- ease with which the machines were connected to the network
ufactured. Additionally, the Job agent will check whether the to be able to receive jobs. The use of web technologies as the
Machine agent represents an AM machine that is compatible backbone of the Living Lab setup meant that the Machine agent
with the gcode held by the Job agent. If these conditions are could be easily and seamlessly setup on a range of hardware.
660 Lorenzo Giunta et al. / Procedia CIRP 119 (2023) 656–661

Fig. 3 shows the demand profile received by the Living Lab However, as previously mentioned, and highlighted in Fig. 3,
during the 3h:30m study, each point on the graphic highlights a the first hour of the study resulted in few prints being accepted
Job agent being accepted for manufacture by a Machine agent. due to technical issues. If the expected number of prints is cal-
A total of 43 jobs were accepted for printing. There is an incon- culated based only on a 2.5 hour timeframe, the maximum num-
sistent demand in the first hour of the study, which was caused ber of prints expected becomes 60.
by technical issues preventing the Job agents from connecting It is possible then to further refine this prediction using the
to the service. These were corrected within the first hour and a findings highlighted in Fig. 4. If the number of prints is cal-
steady state submission of jobs ensued. culated on the basis of a 12-minute timeframe, rather than the
A single researcher was tasked with the removal of jobs from predicted 10, 50 prints are then expected over the 2.5 hours of
the machines. As part of this task, the job had to be removed the study. This is more consistent with the actual number of
from the print bed, the machine interface cleared to mark the prints generated by the ABMS deployed in the Living Lab and
print as removed, and the web interface accessed to inform the demonstrates the utility of using empirical studies to tune our
Machine agent that all this had occurred so that it could accept agent-based numerical models.
new jobs. No filament changes were required during the ses- This 2 minute discrepancy represents a 20% deviation when
sion. based on the expected print time of 10 minutes for each part.
While this may seem particularly noteworthy, it is necessary to
understand whether the 2 minute deviation is a static constant
or scales with the print time required. Additional exploration
would be required, in particular to understand how this time is
impacted by other variables, such as material changes. Either
0:00 0:30 1:00 1:30 2:00 2:30 3:00 3:30 outcome would have significant impact for the computational
Submission Time [HH:MM] model’s design.
The study reinforces both the value of ABMS as a tool to
Fig. 3. Submission of jobs to the LL over a 3.5-hour period. manage manufacturing and the value of a Living Lab to re-
search. A single operator was able to effectively monitor and
Fig. 4 shows a histogram highlighting the difference in the maintain the output of 4 AM machines as well as adjust the en-
real start times of the jobs and the start times as predicted by tire Living Lab setup when a fault was detected. This finding is
the model against the frequency (Count #) with which this dis- encouraging for the future potential of the Living Lab as a tool
crepancy occurs. The result is a normal distribution indicating for research and data collection. It also strengthens the claim
a consistent error around the -2 minute mark. The negative sign that the use of ABMS can aid in improving the responsiveness
indicates the real values are lagging behind those calculated by of manufacturing by highlighting how the need for operator in-
the numerical model. put is reduced when the Job and Machine agents can indepen-
dently handle the job acceptance process.
Additional improvements will be necessary for wider adop-
tion of the ABMS developed for this study, such as added track-
10 ing for the individual machines joining the network rather than
Count [#]

the jobs and agents themselves. This will improve error de-
5
tection and aid operators in responding to issues more rapidly.
However, the results obtained are already encouraging as, bar-
ring the single error with job agents not correctly connecting,
0 the system operated as expected and did not experience capac-
-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 ity issues, allowing for expansion if needed.
Δ Start Time [min] The platform is now in use at the University of Bristol with
students submitting their work for printing. This will provide
Fig. 4. Delta histogram in the LL and numerical model job start times (real - the foundation for a longitudinal study into the usage of an
computed). ABMS that may then be correlated to numerical models and
used to categorize the types of errors, and other forms of “dead
time”, that occur when running a real-world ABMS system.

6. Discussion
7. Conclusion
The study detailed in this paper has highlighted a number
of findings. Firstly, assuming that the “dead time” required for The paper has provided an overview of an ABMS deployed
each print was zero, within the 3.5 hour timeframe for the study at the University of Bristol. The paper discussed the ABMS de-
84 jobs should have been processed (10 minutes required per ployment and how the system was structured. The paper then
job, over 4 machines). This implies only half of the absolute presented a scoping study performed to better understand the
maximum number of jobs was processed. value of ABMS as a tool to manage manufacturing. Further-
Lorenzo Giunta et al. / Procedia CIRP 119 (2023) 656–661 661

more, the paper attempts to understand the potential shortcom- vanced manufacturing technology 67, 1191–1203. doi:10.1007/s00170-
ings in the modelling techniques used to model the behaviour 012-4558-5.
of ABMS. [7] Kuuse, M., 2022. What is distributed manufacturing? URL: https:
//manufacturing-software-blog.mrpeasy.com/distributed-
The findings highlighted by the work performed have evi- manufacturing/.
denced the value of ABMS, allowing a single user to monitor [8] Leitão, P., Restivo, F., 2006. Adacor: A holonic architecture for agile
multiple AM machines. Furthermore, the ABMS did not require and adaptive manufacturing control. Computers in Industry 57, 121–130.
major intervention once in place and functioning, all the deci- doi:10.1016/j.compind.2005.05.005.
sions for the allocation of jobs were taken care of by the agents [9] Lim, M., Zhang, D., 2004. An integrated agent-based approach for respon-
sive control of manufacturing resources. Computers & Industrial Engi-
themselves. neering 46, 221–232. doi:10.1016/j.cie.2003.12.006. special Issue
The software stack presented in this paper to setup the Living on Selected Papers from the 27th. International Conference on Computers
Lab used to run the experiments, made available via an open- and Industrial Engineering, Part 1.
source licence, has proven itself as a successful springboard for [10] Luman, R., Fechner, I., 2022. Trade outlook 2023: slow steaming in rough
future exploration of ABMS. It’s flexibility and ease of deploy- water. URL: https://think.ing.com/downloads/pdf/article/
trade-outlook-slow-steaming-in-rough-waters-what-to-
ment allow for it to be adopted and adapted for future studies. expect-in-2023. [Accessed: 07.03.2023].
The findings from the empirical scoping study provide valu- [11] Ma, A., Frantzén, M., Snider, C., Nassehi, A., 2020. Anarchic manufactur-
able insights into the limitations of current modelling ap- ing: Distributed control for product transition. Journal of Manufacturing
proaches for ABMS. Specifically, the study highlighted the po- Systems 56, 1–10. doi:10.1016/j.jmsy.2020.05.003.
tential issue caused by “dead time”, which occurs when ma- [12] Ma, A., Nassehi, A., Snider, C., 2019. Anarchic manufacturing. Inter-
national Journal of Production Research 57, 2514–2530. doi:10.1080/
chines are idle due to operator input. These insights can in- 00207543.2018.1521534.
form the numerical work previously undertaken by improving [13] Ma, A., Nassehi, A., Snider, C., 2021. Anarchic manufacturing: imple-
predictions for manufacturing output, leading to more accurate menting fully distributed control and planning in assembly. Production
production goals and quotas. Additionally, understanding the & Manufacturing Research 9, 56–80. doi:10.1080/21693277.2021.
causes of “dead time” can inform optimization efforts in the job 1963346.
[14] Mönch, L., Stehli, M., Zimmermann, J., 2003. Fabmas: An agent-based
allocation process and machine agent interactions, ultimately system for production control of semiconductor manufacturing processes,
increasing process efficiency. “Dead time” notwithstanding, the in: Holonic and Multi-Agent Systems for Manufacturing: First Interna-
findings of this empirical study are broadly aligned with those tional Conference on Industrial Applications of Holonic and Multi-Agent
predicted by the numerical model. This reaffirms the value of Systems, HoloMAS 2003, Prague, Czech Republic, September 1-3, 2003.
numerical models for prediction, barring some minor improve- Proceedings 1, Springer. pp. 258–267.
[15] Obi, M., Snider, C., Giunta, L., Goudswaard, M., Gopsill, J., 2022. Coping
ments, as a tool to inform and improve ABMS. with diverse product demand through agent-led type transitions, in: Jezic,
Moving forward, the usability of the ABMS should be fur- G., Chen-Burger, Y.H.J., Kusek, M., Šperka, R., Howlett, R.J., Jain, L.C.
ther explored to evaluate the impact of new functionalities. Ma- (Eds.), Agents and Multi-Agent Systems: Technologies and Applications
terial changes, for example, have not been sufficiently explored. 2022, Springer Nature Singapore, Singapore. pp. 277–286.
The optimisation of models and the calculation of “dead time” [16] Radius, F., 2021. Why distributed manufacturing is the future
of production. URL: https://www.fastradius.com/resources/
when operators have to manually change the material fed into distributed-manufacturing-benefits.
the AM machine could be of great interest for further optimis- [17] Real, R., 2022. Injection moulding plastic lfd cases with re-
ing the job selection logic used by the Machine agents. cycled facemasks. URL: https://dmf-lab.co.uk/blog/
recycling-facemasks-into-covid-test-cases-with-rapid-
prototyping/. [Accessed: 2023.03.08].
[18] v.H. Remko, . Research opportunities for a more resilient post-covid-19
References supply chain – closing the gap between research findings and industry prac-
tice. International Journal of Operations & Production Management 40,
[1] Catapult, T.D., 2018. The rise of distributed autonomous manu- 341–355. doi:10.1108/IJOPM-03-2020-0165.
facturing. URL: https://www.digicatapult.org.uk/wp- [19] Toor, R., 2021. The uk 3d printing filament market in 2021.
content/uploads/2021/11/The_rise_of_distributed_ URL: https://www.filamentive.com/the-uk-3d-printing-
autonomous_manufacturing.pdf. filament-market/.
[2] Duffie, N.A., 1996. Heterarchical control of highly distributed manufactur- [20] Ultimaker, 2021. The uk 3d printing filament market in 2021.
ing systems. International Journal of Computer Integrated Manufacturing URL: https://ultimaker.com/3d-printing-sentiment-index/
9, 270–281. doi:10.1080/095119296131562. overview/uk. [Accessed: 2023.03.08].
[3] Forum, W.E., 2021. Net-zero challenge: The supply chain opportu- [21] Valckenaers, P., Van Brussel, H., Bongaerts, L., Wyns, J., 1997. Holonic
nity. URL: https://www3.weforum.org/docs/WEF_Net_Zero_ manufacturing systems. Integrated Computer-Aided Engineering 4, 191–
Challenge_The_Supply_Chain_Opportunity_2021.pdf. [Ac- 201. doi:10.3233/ICA-1997-4304.
cessed: 07.03.2023]. [22] Van Brussel, H., Wyns, J., Valckenaers, P., Bongaerts, L., Peeters, P., 1998.
[4] Giret, A., Botti, V., 2004. Holons and agents. Journal of intelligent manu- Reference architecture for holonic manufacturing systems: Prosa. Comput-
facturing 15, 645–659. doi:10.1023/B:JIMS.0000037714.56201.a3. ers in Industry 37, 255–274. doi:10.1016/S0166-3615(98)00102-X.
[5] Gopsill, J., Obi, M., Giunta, L., Goudswaard, M., 2022. Queueless: Agent-
based manufacturing for workshop production, in: Jezic, G., Chen-Burger,
Y.H.J., Kusek, M., Šperka, R., Howlett, R.J., Jain, L.C. (Eds.), Agents and
Multi-Agent Systems: Technologies and Applications 2022, Springer Na-
ture Singapore, Singapore. pp. 27–37.
[6] Huang, S.H., Liu, P., Mokasdar, A., Hou, L., 2013. Additive manufacturing
and its societal impact: a literature review. The International journal of ad-

You might also like