An Incremental Model of Lexicon Consensus in a Population of Agents by Means of Grammatical Evolution, Reinforcement Learning and Semantic Rules - Foundations on Natural and Artificial Computation - page 41

was [4]. Other works have been presented to develop a common language or
vocabulary and we will mention only a few ones as [5], [6], [7] or [8]. A good
review about robotic and language evolution is reported in [9]. These methods
and other ones have important advantages but we think grammars have most
expressive power. Another important advantage of grammars is the chance for
expressing a potentially infinitive number of structures with a finite device.
The main aim of this work is to propose an incremental model for creating a
consensual lexicon in a population of artificial agents. We look for this goal by
means of a hybrid algorithm which merge grammatical evolution with semantic
rules and reinforcement learning. In the incremental model the system tries to
evolve a lexicon for a steady set of objects and agents. When the lexicon consen-
sus has emerged new objects and agents can be added to the environment. Then,
the system continues and it tries to find a new consensual lexicon with the new
objects. However, it is very important to maintain the previous knowledge, i.e.
the word-meaning mappings must be preserved because if not the system would
change continuously the lexicon. The experiments show how a language only
with lexicon components can be developed and acquired for a group of agents
and how we can add new objects and agents in order to enhance the vocabulary.
2BcConeps
We need to define some preliminary concepts before to describe the algorithm
and the experiments. A vocabulary or lexicon is an association between a mean-
ing and a symbol. A vocabulary is consensual when all the agents share the same
associations. A meaning can refer to an object or a situation but in this work
we use meanings only as a reference for objects. A symbol or word is a different
way of referring to an object. Both meanings and symbols refer objects and they
define a reference but they do it in a different description level. The idea of
different description levels has been proposed by Deacon [10] or Harnard [11].
In the Deacon and Harnards' proposal a meaning could be a mental represen-
tation, an icon or whatever item we can build as a consequence of the internal
transformation corresponding to the projection of the object in the mind. We
simplify this step in this work and we will implement it in a very simple way as
we will show later. In a second description level, once the physical connection has
been established, the agents can associate each meaning with a symbol. Mapping
between objects and meanings is a complicated problem known as the Symbol
Problem Grounding and we are not going to analyse it deeply in this work. Some
elaborated proposals about how to manage this problem has been reported in
[7], [11] or [12].
In this work we follow a behaviour-based approach (see [13] for details) in the
design of the agents. A simulated agent is endowed with basic behaviours for
acting and sensing. The perception mechanism allows to distinguish objects and
agents. Besides the motor and perceptive behaviours we include two additional
specialized behaviours. The first specialized behaviour is devoted to associate
symbols with meanings and it is triggered when the agent perceives a specific