Modul SARSA pro Modeler neuronových sítí

Abstract

This master thesis describes extension of the program NeuronNetModeler by new module, that consists of reinforcement learning algorithm SARSA and deep reinforcement learning algorithm called Deep SARSA network. Its next goal is parallelization of Deep SARSA learning based on data parallelization. In this thesis are described already mentioned algorithms and their process of learning. Furthermore this thesis contains data parallelization of Deep SARSA network algorithm, that is runned on distributed nodes of supercomputer. Experiments were done with the usage of this implementation, which were focused on their efficiency and speed of learning process.