Reinforcement Lernen mit Regularisierungsnetzwerken

This thesis aims at learning autonomously optimal behavior for high-dimensional control tasks using reinforcement learning with a kernel-based approach. Harnessing the representational power of kernelbased methods we hope to escape the so-called ’curse of dimensionality’, which otherwise implies an exponential growth in the number of basis functions. Specifically, we apply regularization networks as underlying function approximator in least-squares based policy evaluation. The samples used to build this approximation are generated online, from an agent interacting with the environment. This poses an enormous computational challenge since kernel methods inherently scale with O(n3), where n is the number of training samples.