Studies in Learning Automata and Pattern Recognition

点击翻译按钮获取中文摘要

***I propose to undertake various projects that concern the fields of Learning Automata (LA), Anti-Bayesian Pattern Recognition (PR) and Neural Networks (NNs).******LA have been used to model biological systems. They have attracted considerable interest because they can learn the optimal action when operating in stochastic environments. We have pioneered the design of LA that operate in a discretized probability space, and the design of, probably, the fastest LA. We have also used LA to solve numerous real-life hard problems such as optimal web polling, routing for MPLS traffic etc.******Anti-Bayesian PR: In the field of PR, although the moments of the densities of the classes have been used in classification, the properties of the Quantiles have been unused. Recently, we proposed a novel pioneering paradigm named CMQS, Classification by the Moments of Quantile Statistics, which specifically uses these. CMQS is essentially "anti-Bayesian" because it operates in a counter-intuitive manner - it compares the testing sample to a few samples distant from the mean. We have proposed the foundational theory of CMQS for many uni-dimensional and multi-dimensional distributions.******Proposed Work 1. Bayesian Pursuit Algorithms. The fastest family of LA, the Pursuit Algorithms (PAs), utilize estimates of the reward probabilities. These estimates have been of the Maximum Likelihood nature. We propose to investigate families of PAs that involves Bayesian estimates. The work here will primarily be analytic and foundational.******Proposed Work 2. "Elevator"-based LA Solutions. The entire field of LA deals with the model in which the actions that to be learnt are "unordered", i.e., the LA does not have any reason to believe that a1 is better than a2, and so on. However, if the actions can be ordered, they can be considered to be as levels in a building, whereby the learning problem is one of choosing where an elevator must station itself so as to minimize a user-defined criterion. Our research will involve modeling the learning paradigm, and analyzing the unexplored field of generalized "Elevator"-like LA. ******Proposed Work 3. Stochastic Point Location (SPL). Earlier, we had studied the use of LA in the so-called SPL problem and provided discretized, recursive and hierarchical solutions. We propose to design new recursive d-ary search solutions, and to use them in stochastic optimization. ******Proposed Work 4. Applications. We propose to use LA to solve the Goore Game in which the players can also be traitors, to solve the COMMONS game, a generalization of the Prisoner's Dilemma, and in cognitive radios.******Proposed Work 5. PR: We propose to work to develop the theory of Anti-Bayesian PR for non-parametric applications, kernel-based, and unsupervised schemes. We also will work on Chaotic PR.******Proposed Work 6. NN: We propose to develop NNs to enhance the Growing Self-Organizing Map (GSOM) which is a NN clustering algorithm based on Kohonen's Self-Organizing Map.**