Sequential Decision Models Based On Q-Learning Algorithm

Sequential decision problem has been widely used in various fields,while reinforcement learning algorithms provide a powerful tool for solving sequential decision problems.This thesis mainly studies the improvement of Q-learning algorithm,which is a typical algorithm of reinforcement learning,and presents its application into incremental classification problems and group decision making problems.In classification problems,if the classification information used to train the classifier is not obtained at one time but given in the form of sequence,incremental learning provides an important solution.However,the order of samples which are used to train the classifier will seriously affect its performance of the classifier when incremental learning method is employed,especially for the weak classifier.The reason is that the incremental learning method is prone to prematurely add noise data into the training set,reduce the accuracy of the classifier consequently.To solve this problem,this paper proposes an incremental classification model based on Q-learning algorithm.The model uses the classical Q-learning algorithm to select the incremental sample sequence reasonably,therefore reduce the negative impact of noise data,and be able to label the samples automatically during the learning process.At the same time,in order to overcome the computational complexity of the algorithm when the scale of non-labeled training data becomes large,this thesis further proposes a batch incremental classification model.This model reduces the computational complexity of the model and save the storage space effectively as well.Experimental results show that the incremental classification model based on Q-learning algorithm has the advantages of high classification accuracy and strong real-time performance by integrating the incremental learning method and reinforcement learning technology.Multistage group decision-making problem is a kind of typical sequence decision problems.In most applications of the real world,due to the uncertainty of the state space(for example,the state transition probability matrix is completely unknown)faced by the decisionmakers,they need to obtain further information by interacting with the environment dynamically to seek an optimal strategy with higher degree of consensus.Therefore,we establish the basic algorithm model of the multi-stage group decision Q-learning algorithm to learn the optimal group strategy by improving the Q-learning algorithm and improving the iterative process of the algorithm.At the same time,it is proved that the multi-stage group optimal strategy based on Q-learning algorithm is also the strategy with the highest group consensus degree.Finally,an example shows the rationality and feasibility of the algorithm.