We propose a reinforcement learning algorithm, Megh, for live migration of virtual machines that simultaneously reduces the cost of energy consumption and enhances the performance. Megh learns the uncertain dynamics of workloads as-it-goes. Megh uses a dimensionality reduction scheme to project the combinatorially explosive state-action space to a polynomial dimensional space. These schemes enable Megh to be scalable and to work in real-time. We experimentally validate that Megh is more cost-effective and time-efficient than the MadVM and MMT algorithms.
In SIGMOD-2014 blog entry, Guy Lohman asked “Is query optimization a ‘solved’ problem?”. He argued that current query optimizers and their cost models can be critically wrong and can induce significant error in query optimization and database tuning. This motivated us to develop a learning approach to performance tuning of database applications. The objective is to validate the opportunity to devise a tuning strategy that does not need prior knowledge of a cost model. Instead, the cost model is learned through reinforcement learning. We amlgamated techniques of Markov decision process, dimensionality reduction and online learning to develop two methods COREIL and rCOREIL. Theoretical performance analysis and experimental evaluation instantiates their competitive with state-of-the-art adaptive index tuning algorithm, which is dependent on a cost model. Specifically, rCOREIL shows a promise towards emergence of such efficient cost-model oblivious databse tuning systems for modern uncertain and big databases.
Problem of non-rigid registration has become very important in the area of biomedical imaging. In this work, we aimed to improve the graph cuts-based solution to non-rigid registration with a novel data term. This novel data term can efficiently handle the dissimilarities in the intensity patterns between the floating and the reference images which may also arise due to some changes in illumination in addition to motion. We have explored how non-rigid geometry specifically Gromov-Hausdroff distance and Laplace-Beltrami operator can be used in the energy function of the graph cut to further improve the registration accuracy and computational time.
Electroencephalogram (EEG) signals are used to detect the mental states and operation of human brain. Because of EEG signal’s non-stationary and Ergodic nature detection the mental states using this signal is an open area of research. But perfect processing of this signal may help in robotic assistance of physically disabled persons and in building up neuro-motor rehabilitative aids. Our aim was to develop unique methods for filtering, feature extraction and classification of mental states using this and finally to implement that on an embedded system and to control robot arm.