Toward Seamless Transfer from Simulated to Real Worlds: A Dynamically—Rearranging Neural Network Approach

Abstract

In the field of evolutionary robotics artificial neural networks are often used to construct controllers for autonomous agents, because they have useful properties such as the ability to generalize or to be noise-tolerant. Since the process to evolve such controllers in the real-world is very time-consuming, one usually uses simulators to speed up the evolutionary process. By doing so a new problem arises: The controllers evolved in the simulator show not the same fitness as those in the real-world. A gap between the simulated and real environments exists. In order to alleviate this problem we introduce the concept of neuromodulators, which allows to evolve neural networks which can adjust not only the synaptic weights, but also the structure of the neural network by blocking and/or activating synapses or neurons. We apply this concept to a peg-pushing problem for K hepera ™ and compare our method to a conventional one, which evolves directly the synaptic weights. Simulation and real experimental results show that the proposed approach is highly promising.