The Baldwin Effect Revisited: Three Steps Characterized by the Quantitative Evolution of Phenotypic Plasticity Reiji Suzuki and Takaya Arita Graduate School of Human Informatics, Nagoya University Furo-cho, Chikusa-ku, Nagoya 464-8601, Japan {reiji, ari}@info.human.nagoya-u.ac.jp http://www2.create.human.nagoya-u.ac.jp/~{reiji, ari}/ Abstract. An interaction between evolution and learning called the Baldwin effect has been known for a century, but it is still poorly appre- ciated. This paper reports on a computational approach focusing on the quantitative evolution of phenotypic plasticity in complex environment so as to investigate its benefit and cost. For this purpose, we investi- gate the evolution of connection weights in a neural network under the assumption of epistatic interactions. Phenotypic plasticity is introduced into our model, in which whether each connection weight is plastic or not is genetically defined and connection weights with plasticity can be adjusted by learning. The simulation results have clearly shown that the evolutionary scenario consists of three steps characterized by transitions of the phenotypic plasticity and phenotypic variation, in contrast with the standard interpretation of the Baldwin effect that consists of two steps. We also conceptualize this evolutionary scenario by using a hill- climbing image of a population on a fitness landscape. 1 Introduction The Baldwin effect[1] is known as one of the interactions between evolution and learning, which suggests that individual lifetime learning (phenotypic plastic- ity) can influence the course of evolution without the Lamarckian mechanism. This effect explains these interactions by paying attention to balances between benefit and cost of learning through the following two steps[2]. In the first step, lifetime learning gives individual agents chances to change their phenotypes. If the learned traits are useful for agents and make their fitness increase, they will spread in the next population. The learning 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, because of the cost of learning. This step is known as genetic assimilation[3]. Through these steps, learning can guide the genetic acquisition of learned traits without the Lamarckian mechanism in general. Figure 1 roughly shows the concept of the Baldwin effect which consists of two steps described above.