A Distributed Architecture for Real-time Dialogue and On-task Learning of Efficient Co-operative Turn-taking - Coverbal Synchrony in Human-Machine Interaction - page 299

waves (Schaffner, 2006). A way to address the problem of building
more complete models of dialogue is thus to take an interdisciplinary
approach, bringing results from a number of sources to the table at
various levels of abstraction and detail. This is essentially our
approach.
When dealing with the modeling of complex phenomena, building
architectures for systems that integrate multimodal data and exhibit
heterogeneous real-time behaviors, it seems sensible to try to constrain
the possible design space from the outset. One powerful way to do
this is to build multilevel representations (cf. Schwabacher and Gelsey,
1996; Gaud et al., 2007; Dayan, 2000; Arbib, 1987); this may, in fact,
be the only way to get our models right when trying to understand
complex systems such as natural human dialogue. The thrust of this
argument is not that multiple levels are “valid” or even “important”,
as that is a commonly accepted view in science and philosophy,
but, rather, that to map correctly to the many ways sub-systems
interact in such systems they are a critical necessity —that, unless
our simulations are built at fairly high levels of fidelity, we cannot
expect manipulations (expansions, modifications) by its designers to
the architecture at various levels of detail to produce valid results.
Modularity in the architecture is thus highly desirable as it brings
transparency and openness to the architecture, making the modelling
of a highly complex system tractable. However, gross modularity
does not allow the kind of fine-grain representation that we argue is
important for such systems. One drawback of fine-grain modularity is
that decoupling components results in essence in a more distributed
architecture, which calls for non-centralized control schemes. The kind
of modularity and methodology one adopts is critical to the success
of such decoupling.
Many of the existing methodologies that have been offered in
the area of distributed agent-based system construction (cf. Wood
and Deloach, 2000; Wooldridge et al., 2000) suffer from lack of actual
use-case experiences, especially for artificial intelligence projects
that involve construction of single-mind systems. We have built our
present model using the Constructionist Design Methodology (CDM)
(Thórisson et al., 2004) which helps us create complex multi-component
systems at a fairly high level of fidelity, without losing control of the
development process. CDM proposes nine iterative principles to help
with the creation of such systems and has already been applied in the
construction of several systems, both for robots and virtual agents
(cf. Thórisson et al., 2004; Ng-Thow-Hing et al., 2007; Thórisson and
Jonsdottir, 2008). CDM assumes a relatively manual construction
process whereby a large number of pieces are integrated, for example