Language evolution and the Baldwin effect

Yusuke Watanabe, Reiji Suzuki, Takaya Arita
2008 Artificial Life and Robotics  
Recently, a new constructive approach characterized by the use of computational models for simulating the evolution of language has emerged. This paper investigates the interaction between the two adaptation processes in different time scales, evolution and learning of language, by using a computational model. Simulation results show that the fitness increases rapidly and remains at a high level, while the phenotypic plasticity increases together with the fitness but then decreases and
more » ... converges to a medium value. This is regarded as the two-step transition of the so-called Baldwin effect. We investigate the evolutionary dynamics governing the effect. Key words: language evolution, Baldwin effect, genetic algorithm, recurrent neural network, artificial life. If the learned traits are useful for agents and make their fitness increase, they will spread in the next population. The learning behavior acts as a benefit in this step. In the second step, if the environment is sufficiently stable, the evolutionary path finds innate traits that can replace learned traits (genetic assimilation), because of the cost of learning. Through these steps, learning can accelerate the genetic acquisition of learned traits without the Lamarckian mechanism, which has been clearly demonstrated with a variety of models [3] . When analyzing the interaction between evolution and learning, one of the most important aspects is the cost of learning, because the second step of the Baldwin effect can not occur, if learning is ideal, in other words, there is no cost at all arising from the learning process. We adopt a speaker-hearer model proposed by Batali [4], in which each agent used a simple recurrent neural network and structured utterance, in other words, partially compositional communication could be obtained by learning from each other. We use the model in a combined framework of cultural learning and genetic evolution. Adopted cultural learning is an extended version of Iterated Learning proposed by Kirby and Hurford [5], which is based on vertical (oblique) communication from adults to children and horizontal communication between adults. Evolution of the weights in the neural network is achieved by a genetic algorithm. In order to examine whether and how the Baldwin effect might occur, we use a mechanism for the evolution of the plasticity (learnability) of each weight in the neural network as we 3 did in [6] . 7
doi:10.1007/s10015-007-0443-y fatcat:njpc2wsopjdc3hrqptfxnujbta