Improvement And Applications For Q-learning Reinforcement Learning Algorithms

Because reinforcement learning does not teacher signal and keeps improving its cognitive skill through the interaction with environment, it is more perspective for solving complicate optimal and decision-making problems. The paper choose classic Q-Learning in reinforcement learning algorithm and combines different control objectives as experiment model to improve existed reinforcement learning algorithm. Make states in Q-leaming algorithm fuzzy. Moreover, combined with neural network, new understanding and study model are brought out, which are applied in the examples of puzzle, inverted pendulum system, neutralization reaction control and the elevator group control system. The main research results is as fellows:The first part introduces linear 1-stage inverted pendulum and puts forward that the control strategy of Q-leaming algorithm for the inverted pendulum. The accuracy of the control is worse, since the learning system has only four discrete control action. Due to above shortcomings, fuzzy reinforcement learning algorithm is adopted that make outputs of controller continuous to enhance accuracy.The second part is that combination Q-learning and multi-steps Q-learning algorithm is applied in search of the optimal puzzle path, and analyzing and comparing their parameters.The third part is that application of reinforcement learning algorithm in neutralisation processes control is an example, and reinforcement learning algorithm penetrates into biology and chemical field that offer a new study direction for this field.Firstly, its concept was introduced and summarized existed elevator group control strategy method. Secondly, neural network and Q-learning are combined and the combination is applied in the problem of the elevator group control strategy.