Q regression neural network

Q Learning was an important novelty for reinforcement learning. However Q learning becomes inapplicable due to the gigantic size of solution spaces of real world problems. Therefore, an efficient regression method is required not only to generalize the state, action space but also accelerate speed of the learning algorithm. In this thesis, Q regression neural network suggested is a novel regression method obtained as a result of a GRNN form structure used for generalization Q-value function. QRNN works with data generated by Q-Agent. It is tested with popular reinforcement learning benchmarks and its performance is compared with that of both Q-Learning and the other regression methods. Test results show that, QRNN learning efficiency is much higher than the compared methods. Key Words: Reinforcement Learning, Q Learning, Q value function generalization, GRNN