Evolutionary robotics experiments on the emergence of communication

Davide Marocco (1), Angelo Cangelosi (2)

(1) Università della Calabria, (2) University of Plymouth

d.marocco@unical.it

In the computational study of language evolution, new approaches are
sought that use more plausible and realistic cognitive models.  These
should also take data on human and animal communication into
consideration (Tomasello, 2002).  Cognitive robotics uses embodied and
situated robots where communication results from the dynamical
interaction between its physical body, the nervous and cognitive system
and the external physical and social environment (Beer, 1995; Steels,
2003).  Recently, evolutionary robotics (Nolfi & Floreano, 2002; Marocco
et al., 2003) has been proposed to model language evolution.  In this
paper, new experiments are presented based on a recent model of the
emergence of sensorimotor categorization (Nolfi & Marocco, 2002). 
Agents use proprioceptive information to actively explore the
environment (using a three-segment arm) and build sensorimotor
categories of object interactions (e.g.  touch a sphere, avoid a cube). 
In the new simulations, robots also share a lexicon to communicate about
the objects in the environment.  The environment consists of an open
three-dimensional space.  The arm and its interaction with objects are
modelled by means of a rigid body simulator called Vortex.  Initially,
the evolutionary robotics model was used to run a series of experiments
on the role of various social and evolutionary variables in the
emergence of shared communication.  The first independent variable of
the experimental design is the selection of speakers: each agent
receives communication signals solely from its own parent or from any
individual of the population.  This looks at the role of different
social groups of speakers in facilitating shared communication.  The
second independent variable is the time period in which communication is
allowed: agents can communicate right from the initial random generation
or only after the pre-evolution of the ability to touch/avoid the two
objects.  Through this variable it will be possible to investigate the
initial behavioural and cognitive abilities necessary to evolve
communication.  The simulation results show that populations evolve
stable shared communication (i.e.  using two different signals are
produced for the two different objects) mostly when the parents act as
speakers and when signalling is introduced in the second stage (Marocco
et al., 2003).  Additional analyses of results support the following
findings: (a) the emergence of signalling brings direct benefits to the
agents and the population, in terms of increased behavioural skill and
comprehension ability (but the agents' fitness does not assess the
ability to communicate well); (b) there is a benefit in direct
communication between parents and children, not only because of kinship
mechanisms, but also because parents produce more stable and reliable
input signals; (c) the pre-evolution of good sensorimotor and cognitive
abilities permits the establishment of a link between production and
comprehension abilities, especially in the early generations when
signalling is introduced.  This model was extended to study the
emergence of structured lexicons.  Some simulations have looked at the
emergence of different types of signals referring to names of objects
and actions.  These can be related to the linguistic categories of nouns
and verbs in a broad sense in which nouns and verbs could be viewed as
mental categories to represent objects and actions within the process of
interpretation of the world (Pinker, 1994).  In addition, this kind of
approach can give insights on the fact that the meaning of a word is not
simply related to the real object to which the word refers, but more
basically to the way in which the word affects the behavior of the
hearer (Wittgenstein, 1953).  This supports the idea that language could
arise in social negotiation of practical and strictly finalized
situations.  Further extensions of this robotic model include the
ontogenetic learning of language, the use of multi-word utterances, and
the integration of vision and language.  New simulations will focus on
the evolutionary transition from simple communication systems to
syntactic languages.