domingo, 20 de febrero de 2011

A computational model for social learning of language



We present the computational model of social learning of language by Paul Vogt (Tilburg University) and Evert Haasdijk (Vrije Universiteit, Amsterdam). The authors develop the model as part of the NEW TIES project, that is, a simulation platform in which a cultural society could evolve through evolution, individual learning, and social learning. Vogt and Haasdijk (2010) use the platform to set up a two dimensional grid world environment containing edible and inedible plants. If agents eat the edible plants, they gain a fixed amount of energy, and if they eat the inedible plants, they lose a fixed amount of energy. Agents have a visual system detecting their immediate surroundings. Social learning of language considers a population of agents developing a language from scratch. In such model, agents are given an interaction protocol allowing them to exchange expressions, invent new expressions and acquire them from other agents. By means of the local interactions and learning mechanisms, a common language emerges through self-organization, generating a cultural evolution of language.
The lexicon is implemented as an association matrix that maintains frequencies of words and meanings. Agents communicate about paths through their controllers. When an agent hears asn expression, it will try to interpret this expression. If the expression is accompanied by a pointing gesture, the interpreted meaning is only accepted if it is in the learning context.
In the experiments described by the authors, the agents find themselves in an environment where there are two types of plants, nutritious ones and poisonous ones. Agents can distinguish between the two types, but they do not know a priori that one kind is poisonous. Vogt and Haasdijk ran a series of experiments with a population consisting of two kinds of agents: knowers and students. The knowers have controllers that allow them to tackle the problem. The students have a partially randomly constructed controller. When there were no teachers, the number of words exchanged was about 18,000, whereas this number rising to around 76,000 when there were 125 teachers. The rules representing the meanings of words are opaque, because they are not directly accessible to other agents but the meanings are transmitted by means of joint attention. This shows that pedagogy learning (Gergely and Csibra, 2006) can work for different types of knowledge and that scaffolding social learning reduces the cost of learning opaque knowledge, contributing to cumulative cultural evolution.