We examine an evolutionary naming-game model where communicating agents are equipped with an evolutionarily selected learning ability. Such a coupling of biological and linguistic ingredients results in an abrupt transition: upon a small change of a model control parameter a poorly communicating group of linguistically unskilled agents transforms into almost perfectly communicating group with large learning abilities. When learning ability is kept fixed, the transition appears to be continuous. Genetic imprinting of the learning abilities proceeds via Baldwin effect: initially unskilled communicating agents learn a language and that creates a niche in which there is an evolutionary pressure for the increase of learning ability.Our model suggests that when linguistic (or cultural) processes became intensive enough, a transition took place where both linguistic performance and biological endowment of our species experienced an abrupt change that perhaps triggered the rapid expansion of human civilization.
Recently, the emergence and evolution of language attracts a growing interest. In this interdisciplinary field problems like differentiation of languages, development of the speech apparatus or formation of linguistically connected social groups require joint efforts of many specialists such as linguists, neuroscientists, or anthropologists. However, there are also more general questions concerning this almost exclusively human trait. Why do we use words and then combine them into sentences? Why all languages have grammar? To what extent is our brain adapted for acquisition of language? Can learning direct the evolution? This sample questions justify an increasing involvement of researchers also from other disciplines such as artificial intelligence, computer sciences, evolutionary biology or physics [1].
Computer modeling is a frequently used tool in the studies of language evolution. In this technique two main approaches can be distinguished. In the first one, known as an iterated learning model, one is mainly concerned with the transmission of language between successive generations of agents [2,3]. The important issue that the iterated learning model has successfully addressed is the transition from holistic to compositional language. However, since the number of communicating agents is typically very small, the problem of the emergence of linguistic coherence must be neglected in this approach. To tackle this problem Steels introduced a naming game model [4]. In this approach one examines a population of agents trying to establish a common vocabulary for a certain number of objects present in their environment. The change of generations is not required in the naming game-model since the emergence of a common vocabulary is a consequence of the communication processes between agents.
It seems that the iterated learning model and the naming-game model are at two extremes: the first one emphasizes the generational turnover while the latter concentrates on the single-generation (cultural) interactions. Since in the language evolution both aspects are present, it is desirable to examine models that combine evolutionary and cultural processes. In the present paper we introduce such a model. Agents in our model try to establish a common vocabulary like in the naming-game model, but in addition they can breed, mutate, and die. Moreover, they are equipped with an evolutionary trait: learning ability. As a result evolutionary and cultural (learning from peers) processes mutually influence each other. When communication between agents is sufficiently frequent, cultural processes create a favourable niche in which a larger learning ability becomes advantageous. But gradually increasing learning abilities in turn speed up the cultural processes. As a result the model undergoes an abrupt bio-linguistic transition. One can speculate that the proposed model suggests that linguistic and biological processes at a certain point of human history after crossing a certain threshold started to have a strong influence on each other and that resulted in an explosive development of our species. That learning in our model modifies the fitness landscape of a given agent and facilitates the genetic accommodation of learning ability is actually a manifestation of the much debated Baldwin effect [5,6].
In our model we consider a set of agents located at sites of the square lattice of the linear size L. Agents are trying to establish a common vocabulary on a single object present in their environment. An assumption that agents communicate only on a single object does not seem to restrict the generality of our considerations and was already used in some other studies of naming-game [7,8] or language-change [9] models. A randomly selected agent takes the role of a speaker that communicates a word chosen from its inventory to a hearer that is randomly selected among nearest neighbours of the speaker. The hearer tries to recognize the communicated word, namely it checks whether it has it in its inventory. A positive or negative result translates into communicative success or failure, respectively. In some versions of the naming-game model [7,8] success means that both agents retain in their inventories only the chosen word while in the case of failure the hearer adds the communicated word to its inventory.
To implement the learning ability we modified this rule and assigned weights w i (w i > 0) to each i-th word in the inventory. The speaker selects then the i-th word with the probability w i / j w j where summation is over all words in its inventory (if its inventory is empty, it creates a word randomly). If the hearer has the word in its inventory, it is recognized. In addition, each agent k is characterized by its learning ability l k (0 < l k < 1) that is used to modify weights. Namely, in the case of success both speaker and hearer increase the weights of the communicated word by learning abilities of the speaker and hearer, res
This content is AI-processed based on open access ArXiv data.