Adaptive and economic data representation in control architectures of autonomous real-world robots

Abstract

Learning algorithms for autonomous robots in complex, real-world environments usually have to deal with many degrees of freedom and continuous state spaces. Reinforcement learning is a promising concept for unsupervised learning, but common algorithms suffer from huge storage and calculation requirements if they are used to construct an internal model by estimating a value-function for every action in every possible state. In our attempt to approximate this function at the lowest cost, we introduce a flexible method that focuses on the states of greatest interest, and interpolates between them with a fast and easy-to-implement algorithm. In order to provide the highest accuracy to any predefined limit, we enhanced this algorithm by a fast converging multilayer error approximator.