Selection of Neural Architecture and the Environment Complexity Genci Capi 1 , Eiji Uchibe 1,2 and Kenji Doya 1,2 1 CREST, Japan Science and Technology Coorporation 2 ATR, Human Information Science Laboratories “Keihanna Science City”, Kyoto, 619-0288, Japan E-mail:gcapi@atr.co.jp Abstract. In this paper we consider how the complexity of evolved neural controllers depends on the environment using foraging behavior of the Cyber Rodent in two different environments. In the first environment, each fruit can be seen from limited directions and different groups of fruits become ripe in different periods. In the second environment, fruits inside a zone are rewarding and those outside are aversive. After evolution, agents with recurrent neural controller outperformed those with feed-forward controllers by effectively using the memory of border passage. Simulation and experimental results with the Cyber Rodent robot confirmed the selection of appropriate complexity of neural controller, both in size and structure, through evolution. 1 Introduction The neural architectures found in different animal species, at different stages of development, have tremendous diversity. While the sensory-motor mapping and pattern generarion networks are exactly identified in some invertebrates, simple movement by a human is dependent on the complex hierarchical networks of the spinal cord, the brain stem, the cerebellum, the basal ganglia, and the cerebral cortex. What is the origin of such diversity in the network architecture? It is intuitively expected that the more complex the environment, the more complex neural architecture the animals therein should have. Conversely, animals that have more complex neural architecture can perform more challenging behaviors needed for survival and reproduction. The goal of this study is to test such a hypothesis in a evolutionary experiments of simulation and hardware experiments of Cyber Rodent. We focus on the foraging behavior in two different environments. This behavior underlies many more complex behaviors such as following a target or another agent, object manipulation. In the first environment the fruits are randomly scattered and they can not be seen from all directions. Also, they are ripe only for a short time. In this way the agent must combine exploration and directed approach the environment to efficiently eat fruits. In the second environment some fruits are inside and some outside a surrounded zone. The agents get a positive reward if it eats the fruits inside the zone and a negative reward if the fruit eaten is outside. All the fruits have the same characteristics (color and shape). Therefore, the agent must remember if it is inside or outside the zone. We compare the results for a feed-forward Neural Network (FFNN) controller and a recurrent Neural Network (RNN) with two memory units. The simulation and experimental results show that the FFNN controller works well with the first environment but the performance in the second environment is poor. On the other hand the RNN performance in the second