=Paper=
{{Paper
|id=Vol-1855/EUCognition_2016_Part10
|storemode=property
|title=Two Ways (Not) To Design a Cognitive Architecture
|pdfUrl=https://ceur-ws.org/Vol-1855/EUCognition_2016_Part10.pdf
|volume=Vol-1855
|authors=David Vernon
|dblpUrl=https://dblp.org/rec/conf/eucognition/Vernon16
}}
==Two Ways (Not) To Design a Cognitive Architecture==
https://ceur-ws.org/Vol-1855/EUCognition_2016_Part10.pdf
Two Ways (Not) To Design a Cognitive Architecture
David Vernon†
Carnegie Mellon University Africa
Rwanda
Email: vernon@cmu.edu
Abstract—In this short paper, we argue that there are two It continues today under the banner of artificial general in-
conflicting agendas at play in the design of cognitive architectures. telligence, emphasizing human-level intelligence. The term
One is principled: to create a model of cognition and gain an cognitive architecture is employed in a slightly different way in
understanding of cognitive processes. The other is practical: to
build useful systems that have a cognitive ability and thereby the emergent paradigm of cognitive science where it is used
provide robust adaptive behaviour that can anticipate events to denote the framework that facilitates the development of
and the need for action. The first is concerned with advancing a cognitive agent from a primitive state to a fully cognitive
science, the second is concerned with effective engineering. The state. It is a way of dealing with the intrinsic complexity
main point we wish to make is that these two agendas are not of a cognitive system by providing a structure within which
necessarily complementary in the sense that success with one
agenda may not necessarily lead, in the short term at least, to to embed the mechanisms for perception, action, adaptation,
useful insights that lead to success with the other agenda. anticipation, and motivation that enable development over the
systems life-time. Nevertheless, even this slightly different
I. I NTRODUCTION usage reflects an endeavour to construct a viable model that
There are two aspects to the goal of building a cognitive sheds light on the natural phenomenon of cognition.
robot [1]. One is to gain a better understanding of cognition From these perspectives - cognitivist and emergent - a
in general — the so-called synthetic methodology — and the cognitive architecture is an abstract meta-theory of cognition
other is to build systems that have capabilities that are rarely and, as such, focusses on generality and completeness (e.g.
found in technical artifacts (i.e. artificial systems) but are see [3]). It reflects Krichmar’s first aspect of the goal of
commonly found in humans and some animals. The motivation building a cognitive robot: to gain a better understanding
for the first is a principled one, the motivation for the second of cognition in general [1]. We draw from many sources in
is a practical one. Which of these two aspects you choose to shaping these architectures. They are often encapsulated in
focus on has far-reaching effects on the approach you will lists of desirable features (sometimes referred to as desiderata)
end up taking in designing a cognitive architecture. One is or design principles [4], [5], [6], [7]. A cognitive architecture
about advancing science and the other is more about effective schema is not a cognitive architecture: it is a blueprint for the
engineering. These two views are obviously different but they design of a cognitive architecture, setting out the component
are not necessarily complementary. There is no guarantee that functionality and mechanisms for specifying behaviour. It
success in designing a practical cognitive architecture for an describes a cognitive architecture at a level of abstraction
application-oriented cognitive robot will shed any light on the that is independent of the specific application niche that the
more general issues of cognitive science and it is not evident architecture targets. It defines the necessary and sufficient
that efforts to date to design general cognitive architectures software components and their organization for a complete
have been tremendously successful for practical applications. cognitive system. The schema is then instantiated as a cog-
The origins of cognitive architectures reflects the former nitive architecture in a particular environmental niche. This,
principled synthetic methodology. In fact, the term cognitive then, is the first approach to designing a cognitive architecture
architecture can be traced to pioneering research in cognitivist (or a cognitive architecture schema). We refer to it as design
cognitive science by Allen Newell and his colleagues in their by desiderata.
work on unified theories of cognition [2]. As such, a cognitive The second approach is more prosaic, focussing on the
architecture represents any attempt to create a theory that practical necessities of the cognitive architecture and designing
addresses a broad range of cognitive issues, such as attention, on the basis of user requirements. We refer to this as design
memory, problem solving, decision making, and learning, by use case. Here, the goal is to create an architecture that
covering these issues from several aspects including psychol- addresses the needs of an application without being concerned
ogy, neuroscience, and computer science, among others. A whether or not it is a faithful model of cognition. In this
cognitive architecture is, therefore, from this perspective at sense, it is effectively a conventional system architecture,
least, an over-arching theory (or model) of human cognition. rather than a cognitive architecture per se, but one where
† Much of the work described in this paper was conducted while the author
the system exhibits the required attributes and functionality,
was at the University of Skövde, Sweden. This research was funded by the typically the ability to autonomously perceive, to anticipate
European Commission under grant agreement No: 688441, RockEU2. the need for actions and the outcome of those actions, and
Proceedings of EUCognition 2016 - "Cognitive Robot Architectures" - CEUR-WS 42
robot research platform, and the DREAM system architecture
Procedural
Memory
Affective
State
Action
Selection
with its cognitive controller (2 ) [10], [11] which was designed
by use case [12] for use in Robot-Enhanced Therapy targetted
Episodic
Memory Loco-
at children with autism spectrum disorder. The former com-
motion
prises components that reflect generic properties of a cognitive
Attention
Selection
Gaze
Control
iCub
Interface
system; the latter comprises several functional components
Reach
that directly target the needs of therapists who can control
Egosphere
& Grasp
the cognitive architecture through a GUI.
A Priori
Feature
Values
II. C ONCLUSION
Exogenous Endogenous
Vergence
There are two ways not to design a cognitive architecture.
Salience Salience
If your focus is on creating a practical cognitive architecture
for a specific application, you should probably not try to do
iCub
Interface
so by attempting to instantiate a design guided by desiderata;
you are probably better off proceeding in a conventional
manner by designing a system architecture that is driven
Fig. 1. The iCub cognitive architecture (from [9]). by user requirements, drawing on the available repertoire
of AI and cognitive systems algorithms and data-structures.
Conversely, if your focus is a unified theory of cognition —
cognitivist or emergent — then you should probably not try
to do so by developing use-cases and designing a matching
system architecture. You are likely to miss some of the key
considerations that make natural cognitive systems so flexible
and adaptable, and it is unlikely that you will shed much light
on the bigger questions of cognitive science.
R EFERENCES
[1] J. L. Krichmar, “Design principles for biologically inspired cognitive
Fig. 2. Project DREAMs cognitive architecture (from [11]). architectures,” Biologically Inspired Cognitive Architectures, vol. 1, pp.
73–81, 2012.
[2] A. Newell, Unified Theories of Cognition. Cambridge MA: Harvard
University Press, 1990.
to act, learn, and adapt. In this case, the design principles, [3] J. E. Laird, C. Lebiere, and P. S. Rosenbloom, “A standard model of the
or desiderata, do not drive the cognitive architecture — the mind: Toward a common computational framework across artificial in-
requirements do that — but it helps to be aware of them so telligence, cognitive science, neuroscience, and robotics.” AI Magazine,
vol. In Press, 2017.
that you know what capabilities are potentially available and [4] R. Sun, “Desiderata for cognitive architectures,” Philosophical Psychol-
might be deployed to good effect. Significantly, design by use ogy, vol. 17, no. 3, pp. 341–373, 2004.
case implies that it is not feasible to proceed by developing [5] J. L. Krichmar and G. M. Edelman, “Principles underlying the construc-
tion of brain-based devices,” in Proceedings of AISB ’06 - Adaptation in
a cognitive architecture schema and then instantiating it as Artificial and Biological Systems, ser. Symposium on Grand Challenge
a specific cognitive architecture because routing the design 5: Architecture of Brain and Mind, T. Kovacs and J. A. R. Marshall,
through the meta-level schema tacitly abstracts away many of Eds., vol. 2. Bristol: University of Bristol, 2006, pp. 37–42.
[6] J. E. Laird, “Towards cognitive robotics,” in Proceedings of the SPIE
the particularities of the application that makes this approach — Unmanned Systems Technology XI, G. R. Gerhart, D. W. Gage, and
useful. C. M. Shoemaker, Eds., vol. 7332, 2009, pp. 73 320Z–73 320Z–11.
We can recast the distinction between the two motivations [7] D. Vernon, C. von Hofsten, and L. Fadiga, “Desiderata for develop-
mental cognitive architectures,” Biologically Inspired Cognitive Archi-
for building cognitive robots and designing cognitive archi- tectures, vol. 18, pp. 116–127, 2016.
tectures by asking the following question. Should a cognitive [8] D. Vernon, G. Sandini, and G. Metta, “The iCub cognitive architec-
architecture be a specific or a general framework? This is ture: Interactive development in a humanoid robot,” in Proceedings of
IEEE International Conference on Development and Learning (ICDL),
an important design question because a specific instance of Imperial College, London, 2007.
a cognitive architecture derived from a general schema will [9] D. Vernon, C. von Hofsten, and L. Fadiga, A Roadmap for Cognitive
inherit relevant elements but it may also inherit elements that Development in Humanoid Robots, ser. Cognitive Systems Monographs
(COSMOS). Berlin: Springer, 2010, vol. 11.
are not strictly necessary for the specific application domain. [10] D. Vernon, E. Billing, P. Hemeren, S. Thill, and T. Ziemke, “An
Also, it is possible that it is not sufficient, i.e. that it does architecture-oriented approach to system integration in collaborative
not have all the elements that are necessary for the specific robotics research projects - an experience report,” Journal of Software
Engineering for Robotics, vol. 6, no. 1, pp. 15–32, 2015.
application domain. [11] P. Gomez Esteban, H. Cao, A. De Beir, G. Van De Perre, D. Lefeber, and
To illustrate this argument, consider two architectures that B. Vanderborght, “A multilayer reactive system for robots interacting
with children with autism,” in Proceedings of the Fifth International
were designed in these two different manners: the iCub cog- Symposium on New Frontiers in Human-Robot Interaction, 2016.
nitive architecture (Figure 1 ) [8], [9] which was designed by [12] D. David, “Intervention definition,” vol. DREAM Deliverable D1.1,
desiderata [9], [7] for use in a general-purpose open cognitive 2014.
Proceedings of EUCognition 2016 - "Cognitive Robot Architectures" - CEUR-WS 43