On Reinforcement Learning

Reinforcement learning is an important machine learning method. For gaining maximum return, reinforcement learning system can learn the optimal mapping policy through repeatedly interact with environment. Compared to the other learning technologies, the most remarkable advantage is that it can learn without any prior knowledge about the environment. Especially in unknown environment, reinforcement learning can also keep well adaptability and robustness.In real world, some learning problems have successive state space and action space. Some are difficult to be modeled and some have not only one objectives, etc. At the same time, they require quick learning speed and good adaptability. Therefore, for expanding application field, this thesis made systematically research on reinforcement learning and the primary solutions of some existing problems are presented.The basic idea, may steps and existing problems of reinforcement learning areintroduced based on a classical reinforcement learning algorithm——Q-learning algorithm.The experience storage and exploitation in reinforcement learning are researched. Based on the analytical on three existing experience storage methods, a new combinative method is researched. Following the dynamic character, combinative method has been improved for enhance utilization rate of experience and learning performance.The state space compression technologies are investigated. Considering discrete and successive state space separately, a entropy based reinforcement learning algorithm and an auto-generating neural network function approximator method for reinforcement learning are researched. The former makes use of the interrelated degree between state and learning target to make further compression to state space. The latter makes use of evolutionary algorithm to search fit structure and parameters of function approximator, which will ease decision burden of decision maker. Both of two algorithms can save computation and storage source. Thus, they can improve the efficiency of reinforcement learning.A multi-objective reinforcement learning algorithm is researched. Considering the uncommensurability among objectives, combining preference of decision maker with a fuzzy reasoning system can produce a scale to compare two actions. Thus, this method can guide the learning direction and make the learning speed more quickly.