Fitted Q-iteration by Functional Networks for control problems

In this paper a new offline model-free approximate Q-iteration is proposed. Following the idea of Fitted Q-iteration, we use a computational scheme based on Functional Networks, which have been proved to be a powerful alternative to Neural Networks, because they do not require a large number of training samples. We state a condition for the convergence of the proposed technique and we apply it to three classical control problems, namely, a DC motor, a pendulum swing up, a robotic arm. We present a comparative study to show the approximation capabilities of our method with a relatively small number of training samples.