Paolo Petta

File contents

===========================================================
 What is In an Affective Architecture for Situated Agents?
===========================================================
:at: Workshop of the Workpackage 7 (Emotion in Cognition and Action), HUMAINE
:preferred form: oral presentation
:author: Stefan Rank (OFAI),
         Austrian Research Institute for AI, Vienna, Austria,
         stefan.rank@ofai.at
         (+43) 01 5336112-12
:author: Pablo Lucas dos Anjos (HW),
         Heriot-Watt University, Edinburgh, UK,
         anjos@macs.hw.ac.uk
         (+44) 0131 451 4192
:author: Paolo Petta (OFAI),
         Austrian Research Institute for AI, Vienna, Austria,
         Paolo.Petta@ofai.at
         (+43) 01 5336112-12
:author: Ruth Aylett (HW),
         Heriot-Watt University, Edinburgh, UK,
         ruth@macs.hw.ac.uk

Introduction and Characterisation of Approach
----------------------------------------------

The uses of affective architectures are varied, and although different
applications share common characteristics, they are founded on
different---and not always complementary---conceptual assumptions.
Our contribution is an attempt to provide a conceptualisation of the
role of emotional processes in architectures for situated agents,
focussing on the role for bridging the gap between 'higher-' and
'lower-level' aspects of behaviour coordination. This is the topic of
investigations of element 3 of the exemplar proposed in [D7c]_, that
targets human computer interaction involving agents that employ models
of emotional processes for relating to their world which includes
other agents and users.

In our contribution, to characterise our approach, we not only provide
our working definitions on emotion, but also the defining
characteristics of 'higher and lower levels' in architectures for
situated agents, as well as the requirements for application scenarios
that warrant the use of different levels in control mechanisms and the
need for emotional processes as part of the architectures. Further
terms to be analysed include decision-making; action selection;
behaviour coordination; and behaviour arbitration; and their relation
to other notions prominent in agent architectures. Our contribution
thus presents a survey of different terminologies used to describe
agent architectures and their relevance to behaviour coordination and
related processes.

Emotion
~~~~~~~~

Our approach views emotions as a crucial element for intelligent
agents situated in complex social environments (for the important role
of sociality see element 4 in [D7c]_). Emotion, as described in
cognitive appraisal theories [EllsworthScherer2003]_ [Scherer2001a]_
[Scherer2001b]_ [Ortony2003]_ [Frijda1986]_, is characterised as
interplay of processes (rather than the descriptive characterisations
of emotions in dimensional or categorical models). [D3c]_ provides a
definition of an emotion episode taken from Scherer's appraisal
theory. For the purposes of computational modelling, a concise
working description of our approach to emotion could be:

  Emotions involve continuous processes that evaluate stimulus events
  according to subjective criteria, and, for events of subjective
  relevance, influence continuous activity in a synchronised fashion.

A stimulus event, in this context, is understood as an interaction
between the agent and its world, possibly comprising environmental and
internal changes.

A computational model needs to provide all of the infrastructure of a
complete agent [PfeiferScheier1999]_ that is explicitly and implicitly
implied by descriptions of the appraisal process. The involvement of
multiple continuous processes points to this embedding in a complete
agent architecture: it has to allow the synchronised recruitment of
resources. Further, the definition of emotion in [D3c]_ delineates it
from other affective phenomena, like moods and affective dispositions,
that interact through said continuous processes in a complete agent.

In computational models of emotion, such processes are relevant in
'higher' and 'lower levels' of an architecture and especially to link
them. The separation of higher and lower levels is a helpful
simplification that needs to be re-evaluated regularly; cf. that some
theories of human competencies distinguish two separate systems
(planning and control) even for seemingly low-level motor processes
[Glover2004]_. Characteristics that can be used to differentiate
levels in an agent architecture are the time-scale of operations at
different levels [Gat1997]_, the temporal regularities of
representations used, and the independence from the outside world (as
reflected with the reification of the notion of duration at higher
levels). Operations at a lower or reactive level involve more direct
coupling to an agent's environment, the time-scale of operations
corresponds directly to the temporal resolution of the agent's
interface to its world, and no symbolic representations need to be
involved. Examples of processes on higher levels include planning,
that uses a representation of possible actions, or counterfactual
reasoning, that can detach representations from the current state of
the world. The notion of representation has stirred many controversies
in behaviour-based AI [Brooks1991]_ and cognitive science
[Clark1997]_. Relevant properties of higher- and lower-level
representations are the rate of change and if they refer to internal
or environmental conditions.

Having distinguished higher- and lower-levels, the different
influences of emotional processes on these interacting levels, with
different scopes of competence, involve mechanisms for evaluating
interactions on lower levels to provide meaningful information for
higher levels and vice-versa.  Further capabilities are the regulation
of the information flow between the levels and the recruitment of
already existing abilities on the different levels to effect internal
and external changes.

Scenarios for Behaviour Coordination
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Distinctions between higher- and lower-levels only make sense in the
context of a specific agent and the relation to its world. We try to
capture this context for the description of affective agent
architectures as "scenario descriptions". A scenario for the
application of an agent architecture includes the purpose of the
envisioned system as a whole, the kinds, number, and tasks of agents,
the possible interactions between agents and their world, including
the social interaction between agents (possibly mediated through the
environment). Scenarios for the use of agent control architectures can
range from simple virtual worlds with one task and a very restricted
set of interactions to very complex virtual worlds and robot scenarios
that target human-like competencies.  We argue that the complexity of
a scenario is an indicator for the existence of higher and lower
levels on the one hand and emotional processes on the other. For the
purpose of this paper we want to point out some requirements for
scenarios that warrant different architectural levels and emotional
processes and are relevant to behaviour coordination: situatedness
(including temporal criteria for task achievement), possibly ambiguous
and unreliable sensing and acting, multiple conflicting tasks,
variability of resources and resource usage.

Classical definitions of decision-making are anchored in a symbolic
conceptualisation of intelligent action:

  "Decision making is the process of choosing a preferred option or
  course of action from among a set of alternatives."
  [WilsonKeil1999]_

Affect has varied influences on behaviour coordinaton in humans
[LoewensteinLerner2003]_, ranging from the role of expected emotions
in evaluating options, to the influence of immediate emotions, i.e.,
the emotions experienced at the time of decision making. But human,
and in general situated, activity involves different processes of
control that do not necessarily need conscious decisions between
alternatives. A related problem is action selection [Tyrrell1993]_ or
behaviour selection [Gadanho2003]_. The latter term stresses the fact
that what is selected is not a single and atomic action but the
collection of current continuously running situated activities. In
[Tyrrell1993]_ action selection is defined as the problem of "choosing
at each moment in time the most appropriate action out or a repertoire
of possible actions". In order to stress our conviction that for
situated activity there is no selection at "each moment in time", we
prefer to use the term behaviour coordination (used e.g. in
[AlthausChristensen2003]_) to denote the management of ongoing
activities of an entity at different levels (which subsumes decision
making and action selection).

The full contribution presents our terminology for the concepts
involved in behaviour coordination starting form the fundamental
notion of situatedness, and the related concepts concern and
motivation. The terms operation, behaviour, goal, task, and activity
target the area of an agent's executive. The concept of an agent's
lifeworld is fundamental for the analysis of an agent's abilities in a
specific scenario. Finally the crucial notions appraisal, relational
action tendency, and coping are related to their origin in emotion
theories and possible use in affective agent architectures.

Conclusion
-----------

Further steps of our endeavour to clarify architectural terminology
will include the enlargement of the list of reviewed systems and the
refinement of conceptual clarification regarding higher and lower
levels of cognition and action, environment, and scenario. A special
interest is the relation of a scenario, especially its potential for
emotional interactions, to the needs in terms of agent architectures
and virtual world simulations.

By presenting different systems in terms of the terminology used to
describe their take on the problem of behaviour coordination, we aim
at addressing fundamental issues of a comparative approach towards
design, development and evaluation of affective architectures in
emotion-oriented systems. A further result of this effort is the
instigation of discussions in workpackage 7 of the Network of
Excellence HUMAINE about theoretical foundations and practical issues
of matching theory and practice in order to approach a given scenario
of behaviour coordination

Acknowledgments
----------------

The Austrian Research Institute for Artificial Intelligence is
supported by the Austrian Federal Ministry for Education, Science and
Culture and by the Austrian Federal Ministry for Transport, Innovation
and Technology. This research is carried out within the Network of
Excellence Humaine (Contract No. 507422) that is funded by the
European Union's Sixth Framework Programme with support from the
Austrian Funds for Research and Technology Promotion for Industry (FFF
808818/2970 KA/SA). This publication reflects only the authors' views.
The European Union is not liable for any use that may be made of the
information contained herein.

.. References:

.. [AlthausChristensen2003]
   Althaus P., Christensen H.I.: Behaviour Coordination in Structured Environments,
   Advanced Robotics, 17(7):657-674, 2003.
.. [Brooks1991]
   Brooks R.A.: Intelligence without Representation, Artificial Intelligence,
   Special Volume: Foundations of Artificial Intelligence, 47(1-3), 139-159, 1991.
.. [Clark1997]
   Clark A.: Being There - Putting Brain, Body and World Together Again,
   MIT Press/Bradford Books, Cambridge/London, 1997.
.. [D3c]
   http://emotion-research.net/deliverables/D3c.pdf
   last accessed 20-05-2005
.. [D7c]
   http://emotion-research.net/deliverables/D7c%20potential%20exemplars%20cog_%20action.pdf,
   last accessed 20-05-2005
.. [EllsworthScherer2003]
   Ellsworth P.C., Scherer K.R.: Appraisal Processes in Emotion,
   in Davidson R.J. et al. (eds.), Handbook of Affective Sciences,
   Oxford University Press, Oxford New York, pp.572-595 (Chapter 29), 2003.
.. [Frijda1986]
   Frijda N.H.: The Emotions,
   Cambridge University Press, Editions de la Maison des Sciences de l'Homme, Paris, 1986.
.. [Gadanho2003]
   Gadanho S.C.: Learning Behavior-Selection by Emotions and Cognition in a Multi-Goal Robot Task,
   Journal of Machine Learning Research, 4:385-412, 2003.
.. [Gat1997]
   Gat E.: On Three-Layer Architectures,
   in Kortenkamp D., Bonasso R.P., Murphy R. (eds.): Artificial Intelligence and Mobile Robots,
   MIT/AAAI Press, 1997.
.. [Glover2004]
   Glover S.: Separate visual representations in the planning and control of action,
   Behavioural and Brain Sciences, 27:3-78, 2004.
.. [LoewensteinLerner2003]
   Loewenstein G., Lerner J.S.: The Role of Affect in Decision Making,
   in Davidson R.J. et al. (eds.), Handbook of Affective Sciences,
   Oxford University Press, Oxford New York, pp.619-642 (Chapter 31), 2003.
.. [Ortony2003]
   Ortony A.: On Making Believable Emotional Agents Believable,
   in Trappl R. et al. (eds.), Emotions in Humans and Artifacts,
   MIT Press Cambridge, MA, pp.189-212, 2003.
.. [PfeiferScheier1999]
   Pfeifer R., Scheier C.: Understanding Intelligence,
   MIT Press/Bradford Books, Cambridge/London, 1999.
.. [Scherer2001a]
   Scherer K.R., Schorr A., Johnstone T. (ed.): Appraisal Processes in Emotion:
   Theory, Methods, Research, Oxford University Press, Oxford New York, 2001.
.. [Scherer2001b]
   Scherer K.R.: Appraisal Considered as a Process of Multilevel Sequential Checking,
   in Scherer K.R. et al. (eds.), Appraisal Processes in Emotion: Theory, Methods, Research,
   Oxford University Press, Oxford New York, pp.92-120, 2001.
.. [Tyrrell1993]
   Tyrrell T.: Computational Mechanisms for Action Selection,
   University of Edinburgh, UK, Ph.D. thesis, 1993.
.. [WilsonKeil1999]
   Wilson R.A., Keil F.C. (eds.): The MIT Encyclopedia of the Cognitive Sciences,
   MIT Press, Cambridge/Boston/London, 1999.