Specifying the reward function of a Markov decision process (MDP) can be demanding, requiring human assessment of the precise quality of, and tradeoffs among, various states and actions. However, reward functions often possess considerable structure which can be leveraged to streamline their specification. We develop new, decision-theoretically sound heuristics for eliciting rewards for factored MDPs whose reward functions exhibit additive independence. Since we can often find good policies without complete reward specification, we also develop new (exact and approximate) algorithms for robust optimization ofimprecise-reward MDPs with such additive reward. Our methods are evaluated in two domains: autonomic computing and assistive technology.

It has been known for a long time that visual task, such as reading, counting and searching, greatly influences eye movement patterns. Perhaps the best known demonstration of this is the celebrated study of Yarbus showing that different eye movement trajectories emerge depending on the visual task that the viewers are given. The objective of this paper is to develop an inverse Yarbus process whereby we can infer the visual task by observing the measurements of a viewer’s eye movements while executing the visual task. The method we are proposing is to use Hidden Markov Models (HMMs) to create a probabilistic framework to infer the viewer’s task from eye movements.

Human Behavior Analysis in Uncontrolled Environmentscan be categorized in two main challenges:1) Feature extraction and 2) Behavior analysisfrom a set of corporal language vocabulary. Inthis work, we present our achievements characterizingsome simple behaviors from visual data ondifferent real applications and discuss our plan forfuture work: low level vocabulary definition frombag-of-gesture units and high level modelling andinference of human behaviors.

We consider the problem of learning to act in partially observable, continuous-state-and-action worlds where we have abstract prior knowledge about the structure of the optimal policy in the form of a distribution over policies. Using ideas from planning-as-inference reductions and Bayesian unsupervised learning, we cast Markov Chain Monte Carlo as a stochastic, hill-climbing policy search algorithm. Importantly, this algorithm's search bias is directly tied to the prior and its MCMC proposal kernels, which means we can draw on the full Bayesian toolbox to express the search bias, including nonparametric priors and structured, recursive processes like grammars over action sequences. Furthermore, we can reason about uncertainty in the search bias itself by constructing a hierarchical prior and reasoning about latent variables that determine the abstract structure of the policy. This yields an adaptive search algorithm---our algorithm learns to learn a structured policy efficiently. We show how inference over the latent variables in these policy priors enables intra- and intertask transfer of abstract knowledge. We demonstrate the flexibility of this approach by learning meta search biases, by constructing a nonparametric finite state controller to model memory, by discovering motor primitives using a simple grammar over primitive actions, and by combining all three.

The scalability of algorithms for solving Markov Decision Processes (MDPs) has been a limiting factor for MDPs as a modeling tool. This dissertation develops theoretical and empirical techniques for solving larger MDPs than was possible before, and aims to demonstrate the achieved progress by applying these new algorithms to a real-world problem.

In this proposal, we introduce Bayesian Abductive Logic Programs (BALP), a probabilistic logic that adapts Bayesian Logic Programs (BLPs) for abductive reasoning. Like BLPs, BALPs also combine first-order logic and Bayes nets. However, unlike BLPs, which use deduction to construct Bayes nets, BALPs employ logical abduction. As a result, BALPs are more suited for problems like plan/activity recognition that require abductive reasoning. In order to demonstrate the efficacy of BALPs, we apply it to two abductive reasoning tasks — plan recognition and natural language understanding.

Chinese Pinyin input methods are very important for Chinese language processing. In many cases, users may make typing errors. For example, a user wants to type in "shenme" (什么, meaning "what" in English) but may type in "shenem" instead. Existing Pinyin input methods fail in converting such a Pinyin sequence with errors to the right Chinese words. To solve this problem, we developed an efficient error-tolerant Pinyin input method called "CHIME'' that can handle typing errors. By incorporating state-of-the-art techniques and language-specific features, the method achieves a better performance than state-of-the-art input methods. It can efficiently find relevant words in milliseconds for an input Pinyin sequence.

Constrained partially observable Markov decision processes (CPOMDPs) extend the standard POMDPs by allowing the specification of constraints on some aspects of the policy in addition to the optimality objective for the value function. CPOMDPs have many practical advantages over standard POMDPs since they naturally model problems involving limited resource or multiple objectives. In this paper, we show that the optimal policies in CPOMDPs can be randomized, and present exact and approximate dynamic programming methods for computing randomized optimal policies. While the exact method requires solving a minimax quadratically constrained program (QCP) in each dynamic programming update, the approximate method utilizes the point-based value update with a linear program (LP). We show that the randomized policies are significantly better than the deterministic ones. We also demonstrate that the approximate point-based method is scalable to solve large problems.

Imprecise-reward Markov decision processes (IRMDPs) are MDPs in which the reward function is only partially specified (e.g., by some elicitation process). Recent work using minimax regret to solve IRMDPs has shown, despite their theoretical intractability, how the set of policies that are nondominated w.r.t. reward uncertainty can be exploited to accelerate regret computation. However, the number of nondominated policies is generally so large as to undermine this leverage. In this paper, we show how the quality of the approximation can be improved online by pruning/adding nondominated policies during reward elicitation, while maintaining computational tractability. Drawing insights from the POMDP literature, we also develop a new anytime algorithm for constructing the set of nondominated policies with provable (anytime) error bounds. These bounds can be exploited to great effect in our online approximation scheme.

In multidimensional classification the goal is to assign an instance to a set of different classes. This task is normally addressed either by defining a compound class variable with all the possible combinations of classes (label power-set methods, LPMs) or by building independent classifiers for each class (binary-relevance methods, BRMs). However, LPMs do not scale well and BRMs ignore the dependency relations between classes. We introduce a method for chaining binary Bayesian classifiers that combines the strengths of classifier chains and Bayesian networks for multidimensional classification. The method consists of two phases. In the first phase, a Bayesian network (BN) that represents the dependency relations between the class variables is learned from data. In the second phase, several chain classifiers are built, such that the order of the class variables in the chain is consistent with the class BN. At the end we combine the results of the different generated orders. Our method considers the dependencies between class variables and takes advantage of the conditional independence relations to build simplified models. We perform experiments with a chain of naive Bayes classifiers on different benchmark multidimensional datasets and show that our approach outperforms other state-of-the-art methods.