The Autonomy of Automated Systems: Social Systems and the Multi-level Autonomy

Social Systems and the Multi-level Autonomy
The Autonomy of Automated Systems
Andrea Omicini Giovanni Sartor
European University Institute
Alma Mater Studiorum—Universit`a di Bologna
Autonomous Weapons Systems – Law, Ethics, Policy
European University Institute in Florence, Italy
24 April 2014
Omicini, Sartor (EUI & Univ. Bologna) Social Systems and the Multi-level Autonomy AWS-LEP – 24/4/2014 1 / 22

Outline
1 Agents & Multi-Agent Systems
2 Multi-level Autonomy
3 Conclusion

Agents & Multi-Agent Systems
Outline
3 Conclusion

Agents, Coordination, and Societies I
Complex systems as MAS
Nowadays, most of the complex computational systems of interest
can be thought, modelled, and built as multi-agent systems (MAS)
[Zambonelli and Omicini, 2004]
In a MAS, many autonomous components (the agents) with their
own individual goal interact in order to achieve the overall designed
system goal

Agents, Coordination, and Societies II
Agent societies & coordination
A group of agents coordinating in order to achieve some (local,
non-individual) goal is called an agent society
Agent societies are built around coordination media
[Gelernter and Carriero, 1992], encapsulating social (coordination)
laws
Agent societies may have collective goals, possibly independent of the
individual agent goals [Ciancarini, 1996]
Examples: norms as coordination mechanisms enforced by
computational institutions (aka electronic institutions)
[Noriega and Sierra, 2002] working as the coordination abstractions

A Layered View of MAS I
Layering
From the software engineering viewpoint, agent societies represent a
layering mechanism [Molesini et al., 2006]
There, agents, societies, individual and social goals are conceptual
tools to be used at the most suitable level of abstraction
Each group of agents could be seen in principle as a single agent at a
higher level of abstraction
Viceversa, each agent could be modelled / built as an agent society
at deeper level of detail – with the global MAS level working as the
uppermost layer
Accordingly, individual / social / global goals could in principle be
layered in the same way

A Layered View of MAS II
Autonomous systems out of autonomous components
As a result, a MAS could be in principle conceived, designed, and
built as an autonomous system made of autonomous components
Even more, each agent society could be handled in the same way
So, autonomy could be conceived as a multi-level property of
computational systems designed as MAS, associated to each agent, to
agent societies, and to the global MAS level as well

Multi-level Autonomy
Outline
3 Conclusion

Diverse Levels of Autonomy I
Independence
The key point here is the fact that conceptually there is no direct
dependence between the diverse levels of autonomy at the diﬀerent
levels of the MAS
In the case of coordinated systems, the coordination media could
embed
the reactive behaviour for an automatic coordinated behaviour
the implicit mechanisms for a teleonomic behaviour of the agent society
the operational plans for a teleologic social behaviour—for instance, by
adopting ReSpecT tuple centres [Omicini and Denti, 2001] as the
coordination media
All this, independently of the level of autonomy of the individual
agents composing the agent society

Diverse Levels of Autonomy II
Artefacts: making it more complex
Even more, a more articulated reference model for MAS, such as the
agents & artefacts (A&A) meta-model [Omicini et al., 2008] could
make the picture even more intricate
An A&A MAS is basically composed by agents and artefacts, where
artefacts are the tools that agents use to achieve their own goals
as such, they are typically automated, deterministic entities, which are
not required to be autonomous
So, for instance, by extending layering to include artefacts, any level
of a MAS could then feature any sort of automatic / autonomic /
autonomous behaviour, essentially independently of any other
individual / social / global property

Issues I
Classifying complex MAS
Classifying a complex MAS – either a software or a robotic one – as
automatic / teleonomic / teleologic is not necessarily a trivial task
It may require in general some articulated model
Arguably, a multi-level model of autonomy

Issues II
Global MAS level might not be enough
What if we assume that the global level of autonomy of a MAS is the
most relevant system feature?
Issues such as responsibility and liability cannot a priori be reduced to
the simple observation of the main level of a MAS
Existing works on collective moral and responsibility typically refer to
human groups, and just account for a two-level layering
For instance
an individual may participate to more than one MAS, possibly
expressing diﬀerent levels of autonomy in the diverse contexts
however, being a single component of more than one system, it could
in principle work as an element of inter-system interference, possibly
undetected, which could make issues like responsibility and liability
much more complex

Issues III
Non-determinism
This is particularly the case of non-deterministic systems—for instance in the case
of stochastic behaviours in nature-inspired models
There, in fact, upper-level behaviour could appear by emergence without any linear
connection with the lower level components—as in the case of swarm systems
For instance, coordination media could used to encapsulate local interaction
leading to self-organising behaviours [Ricci et al., 2007]
There, teleonomic components would self-organising around automatic
abstractions such as tuple-based coordination media
The resulting behaviour would be essentially unpredictable
and be classiﬁed as either teleonomic – since it tends to autonomously preserve
some essential system property – or even (possibly) teleologic—when coordination
artefacts would contain the local policies explicitly designed to produce the overall
self-organising behaviour [Omicini et al., 2004]

Conclusion
Outline
3 Conclusion

Conclusion
Artificial Systems. . . I
. . . are not mere social systems
where humans and groups provide essentially two distinct levels of
abstraction
to be used for the attribution of properties
in artificial systems, many levels can be used, and possibly with the
same set of criteria
. . . are not natural systems
where the hierarchical view concerns fundamentally diverse layers
each one with its own entities and laws
in artificial systems, the many levels possibly available should be
described with a uniform set of criteria

Conclusion
Multi-level Autonomy. . .
. . . mandates for new concepts and tools
where artiﬁcial systems of any sorts could be understood and classiﬁed
according to their (possibly diverse) levels of autonomy
along with the many related concepts, such as (collective)
responsibility and liability

Outline
3 Conclusion

Bibliography
Bibliography I
Ciancarini, P. (1996).
Coordination models and languages as software integrators.
ACM Computing Surveys, 28(2):300–302.
Gelernter, D. and Carriero, N. (1992).
Coordination languages and their signiﬁcance.
Communications of the ACM, 35(2):97–107.
Molesini, A., Omicini, A., Ricci, A., and Denti, E. (2006).
Zooming multi-agent systems.
In M¨uller, J. P. and Zambonelli, F., editors, Agent-Oriented Software
Engineering VI, volume 3950 of Lecture Notes in Computer Science,
pages 81–93. Springer Berlin Heidelberg.
6th International Workshop (AOSE 2005), Utrecht, The Netherlands,
25–26 July 2005. Revised and Invited Papers.

Bibliography
Bibliography II
Noriega, P. and Sierra, C. (2002).
Electronic Institutions: Future trends and challenges.
In Klusch, M., Ossowski, S., and Shehory, O., editors, Cooperative
Information Agents VI, volume 2446 of Lecture Notes in Artiﬁcial
Intelligence, pages 14–17. Springer Berlin Heidelberg.
6th International Workshop (CIA 2002), Madrid, Spain,
18–20 September 2002. Proceedings.
Omicini, A. and Denti, E. (2001).
From tuple spaces to tuple centres.
Science of Computer Programming, 41(3):277–294.

Bibliography
Bibliography III
Omicini, A., Ricci, A., and Viroli, M. (2008).
Artifacts in the A&A meta-model for multi-agent systems.
Autonomous Agents and Multi-Agent Systems, 17(3):432–456.
Special Issue on Foundations, Advanced Topics and Industrial
Perspectives of Multi-Agent Systems.
Omicini, A., Ricci, A., Viroli, M., Castelfranchi, C., and Tummolini, L.
(2004).
Coordination artifacts: Environment-based coordination for intelligent
agents.
In Jennings, N. R., Sierra, C., Sonenberg, L., and Tambe, M., editors,
3rd international Joint Conference on Autonomous Agents and
Multiagent Systems (AAMAS 2004), volume 1, pages 286–293, New
York, USA. ACM.

Bibliography
Bibliography IV
Ricci, A., Omicini, A., Viroli, M., Gardelli, L., and Oliva, E. (2007).
Cognitive stigmergy: Towards a framework based on agents and
artifacts.
In Weyns, D., Parunak, H. V. D., and Michel, F., editors,
Environments for MultiAgent Systems III, volume 4389 of LNCS,
pages 124–140. Springer.
3rd International Workshop (E4MAS 2006), Hakodate, Japan,
8 May 2006. Selected Revised and Invited Papers.
Zambonelli, F. and Omicini, A. (2004).
Challenges and research directions in agent-oriented software
engineering.
Autonomous Agents and Multi-Agent Systems, 9(3):253–283.
Special Issue: Challenges for Agent-Based Computing.

Social Systems and the Multi-level Autonomy
The Autonomy of Automated Systems
Andrea Omicini Giovanni Sartor
European University Institute
Alma Mater Studiorum—Universit`a di Bologna
Autonomous Weapons Systems – Law, Ethics, Policy
European University Institute in Florence, Italy
24 April 2014

The Autonomy of Automated Systems: Social Systems and the Multi-level Autonomy

More Related Content

More from Andrea Omicini (20)

Recently uploaded (20)

The Autonomy of Automated Systems: Social Systems and the Multi-level Autonomy