This thesis addresses the open problem of automatically discovering hierarchical structure in reinforcement learning. Current algorithms for reinforcement learning fail to scale as problems become more complex. Many complex environments empirically exhibit hierarchy and can be modeled as interrelated subsystems, each in turn with hierarchic structure. Subsystems are often repetitive in time and space, meaning that they reoccur as components of different tasks or occur multiple times in different circumstances in the environment. A learning agent may sometimes scale to larger problems if it successfully exploits this repetition. Evidence suggests that a bottom up approach that repetitively finds building-blocks at one level of abstraction and uses them as background knowledge at the next level of abstraction, makes learning in many complex environments tractable. An algorithm, called HEXQ, is described that automatically decomposes and solves a multi-dimensional Markov decision problem (MDP) by constructing a multi-level hierarchy of interlinked subtasks without being given the model beforehand. The effectiveness and efficiency of the HEXQ decomposition depends largely on the choice of representation in terms of the variables, their temporal relationship and whether the problem exhibits a type of constrained stochasticity. The algorithm is first developed for stochastic shortest path problems and then extended to infinite horizon problems. The operation of the algorithm is demonstrated using a number of examples including a taxi domain, various navigation tasks, the Towers of Hanoi and a larger sporting problem. The main contributions of the thesis are the automation of (1)decomposition, (2) sub-goal identification, and (3) discovery of hierarchical structure for MDPs with states described by a number of variables or features. It points the way to further scaling opportunities that encompass approximations, partial observability, selective perception, relational representations and planning. The longer term research aim is to train rather than program intelligent agents

Sequential supervised learning problems involve assigning a class label to each item in a sequence. Examples include part-of-speech tagging and text-to-speech mapping. A very general-purpose strategy for solving such problems is to construct a recurrent sliding window (RSW) classifier, which maps some window of the input sequence plus some number of previously-predicted items into a prediction for the next item in the sequence. This paper describes a general purpose implementation of RSW classifiers and discusses the highly practical issue of how to choose the size of the input window and the number of previous predictions to incorporate. Experiments on two real-world domains show that the optimal choices vary from one learning algorithm to another. They also depend on the evaluation criterion (number of correctly-predicted items versus number of correctly-predicted whole sequences). We conclude that window sizes must be chosen by cross-validation. The results have implications for the choice of window sizes for other models including hidden Markov models and conditional random fields. / Graduation date: 2004

Reinforcement Learning (RL) is the study of learning agents that improve their performance from rewards and punishments. Most reinforcement learning methods optimize the discounted total reward received by an agent, while, in many domains, the natural criterion is to optimize the average reward per time step. In this thesis, we introduce a model-based average reward reinforcement learning method called "H-learning" and show that it performs better than other average reward and discounted RL methods in the domain of scheduling a simulated Automatic Guided Vehicle (AGV). We also introduce a version of H-learning which automatically explores the unexplored parts of the state space, while always choosing an apparently best action with respect to the current value function. We show that this "Auto-exploratory H-Learning" performs much better than the original H-learning under many previously studied exploration strategies. To scale H-learning to large state spaces, we extend it to learn action models and reward functions in the form of Bayesian networks, and approximate its value function using local linear regression. We show that both of these extensions are very effective in significantly reducing the space requirement of H-learning, and in making it converge much faster in the AGV scheduling task. Further, Auto-exploratory H-learning synergistically combines with Bayesian network model learning and value function approximation by local linear regression, yielding a highly effective average reward RL algorithm. We believe that the algorithms presented here have the potential to scale to large applications in the context of average reward optimization. / Graduation date:1996

Thesis (Ph.D.)--Brown University, 2005. / Vita. Thesis advisor: Stuart Geman. Includes bibliographical references (leaves 7-9, 31-33, 64-68, 107-107, 131-132, 155-157, 267-268). Also available online.

Thesis (Ph. D.)--Rutgers University, 2009. / "Graduate Program in Biomedical Engineering." Includes bibliographical references (p. 54-55).

Thesis (M. Phil.)--Hong Kong University of Science and Technology, 2003. / Includes bibliographical references (leaves 46-49). Also available in electronic version. Access restricted to campus users.

No description available.

Recent work has shown that AdaBoost can be viewed as an algorithm that maximizes the margin on the training data via functional gradient descent. Under this interpretation, the weight computed by AdaBoost, for each hypothesis generated, can be viewed as a step size parameter in a gradient descent search. Friedman has suggested that shrinking these step sizes could produce improved generalization and reduce overfitting. In a series of experiments, he showed that very small step sizes did indeed reduce overfitting and improve generalization for three variants of Gradient_Boost, his generic functional gradient descent algorithm. For this report, we tested whether reduced learning rates can also improve generalization in AdaBoost. We tested AdaBoost (applied to C4.5 decision trees) with reduced learning rates on 28 benchmark datasets. The results show that reduced learning rates provide no statistically significant improvement on these datasets. We conclude that reduced learning rates cannot be recommended for use with boosted decision trees on datasets similar to these benchmark datasets. / Graduation date: 2002

Thesis (Ph.D.)--Hong Kong University of Science and Technology, 2007. / Includes bibliographical references (leaves 102-108). Also available in electronic version.

Thesis (M.S.)--Oregon State University, 2006. / Printout. Includes bibliographical references (leaves 42-46). Also available on the World Wide Web.