Clustering methods based on reinforcement learning (Master thesis)
Clustering is one of the most popular problems in machine learning and data mining. It belongs to the category of unsupervised learning problems since no label information is provided to assist in partitioning the data points into coherent groups. Although clustering is an unsupervised problem, it is possible to view clustering from a reinforcement learning perspective. In reinforcement learning, an agent learns an action policy that solves a sequential decision problem using reinforcement signals provided by the environment. In reinforcement-based clustering, the clustering system learns through reinforcements to follow the desired clustering policy. The previous method of this type (RGCL algorithm) trains a team of binary stochastic units to perform on-line clustering. Each unit corresponds to a cluster and the weights of each stochastic unit correspond to a representative point (centroid) of the respective cluster. The team of stochastic units is trained to perform clustering using the REINFORCE algorithm by exploiting properly defined reinforcement signals provided by the environment. In this thesis we propose two extensions of the RGCL algorithm based on the use of a single stochastic multinomial unit instead of a team of binary stochastic units. In the first method the unit is trained on-line based on the REINFORCE framework using immediate reinforcement signals. In the second method the stochastic multinomial unit is trained in a batch mode based on the REINFORCE framework using delayed reinforcement signals. In both cases the weight update equations are derived so that the weight updates lead to the stochastic minimization of the well-known k-means clustering error. An experimental study has been conducted using synthetic and real datasets to assess the performance of the proposed methods. The experimental results indicate that improved clustering results are obtained in the majority of cases.
|Institution and School/Department of submitter:||Πανεπιστήμιο Ιωαννίνων. Πολυτεχνική Σχολή. Τμήμα Μηχανικών Ηλεκτρονικών Υπολογιστών και Πληροφορικής|
|Subject classification:||Reinforcement learning|
|Keywords:||Clustering,Reinforcement learning,Machine learning|
|Appears in Collections:||Διατριβές Μεταπτυχιακής Έρευνας (Masters)|
Please use this identifier to cite or link to this item:This item is a favorite for 0 people.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.