We develop a qualitative theory of Markov Decision Processes (MDPs) and Partially Observable MDPs that can be used to model sequential decision making tasks when only qualitative information is available. Our approach is based upon an order-of-magnitude approximation of both probabilities and utilities, similar to epsilon-semantics. The result is a qualitative theory that has close ties with the standard maximum-expected-utility theory and is amenable to general planning techniques.

This paper uses decision-theoretic principles to obtain new insights into the assessment and updating of probabilities. First, a new foundation of Bayesianism is given. It does not require infinite atomless uncertainties as did Savage s classical result, AND can therefore be applied TO ANY finite Bayesian network.It neither requires linear utility AS did de Finetti s classical result, AND r ntherefore allows FOR the empirically AND normatively desirable risk r naversion.Finally, BY identifying AND fixing utility IN an elementary r nmanner, our result can readily be applied TO identify methods OF r nprobability updating.Thus, a decision - theoretic foundation IS given r nto the computationally efficient method OF inductive reasoning r ndeveloped BY Rudolf Carnap.Finally, recent empirical findings ON r nprobability assessments are discussed.It leads TO suggestions FOR r ncorrecting biases IN probability assessments, AND FOR an alternative r nto the Dempster - Shafer belief functions that avoids the reduction TO r ndegeneracy after multiple updatings.r n

We extend the language of influence diagrams to cope with decision scenarios where the order of decisions and observations is not determined. As the ordering of decisions is dependent on the evidence, a step-strategy of such a scenario is a sequence of dependent choices of the next action. A strategy is a step-strategy together with selection functions for decision actions. The structure of a step-strategy can be represented as a DAG with nodes labeled with action variables. We introduce the concept of GS-DAG: a DAG incorporating an optimal step-strategy for any instantiation. We give a method for constructing GS-DAGs, and we show how to use a GS-DAG for determining an optimal strategy. Finally we discuss how analysis of relevant past can be used to reduce the size of the GS-DAG.

This paper investigates value function approximation in the context of zero-sum Markov games, which can be viewed as a generalization of the Markov decision process (MDP) framework to the two-agent case. We generalize error bounds from MDPs to Markov games and describe generalizations of reinforcement learning algorithms to Markov games. We present a generalization of the optimal stopping problem to a two-player simultaneous move Markov game. For this special problem, we provide stronger bounds and can guarantee convergence for LSTD and temporal difference learning with linear value function approximation. We demonstrate the viability of value function approximation for Markov games by using the Least squares policy iteration (LSPI) algorithm to learn good policies for a soccer domain and a flow control problem.

We show that maximum entropy (maxent) models can be modeled with certain kinds of HMMs, allowing us to construct maxent models with hidden variables, hidden state sequences, or other characteristics. The models can be trained using the forward-backward algorithm. While the results are primarily of theoretical interest, unifying apparently unrelated concepts, we also give experimental results for a maxent model with a hidden variable on a word disambiguation task; the model outperforms standard techniques.

We present a principled and efficient planning algorithm for collaborative multiagent dynamical systems. All computation, during both the planning and the execution phases, is distributed among the agents; each agent only needs to model and plan for a small part of the system. Each of these local subsystems is small, but once they are combined they can represent an exponentially larger problem. The subsystems are connected through a subsystem hierarchy. Coordination and communication between the agents is not imposed, but derived directly from the structure of this hierarchy. A globally consistent plan is achieved by a message passing algorithm, where messages correspond to natural local reward functions and are computed by local linear programs; another message passing algorithm allows us to execute the resulting policy. When two portions of the hierarchy share the same structure, our algorithm can reuse plans and messages to speed up computation.

Filtering---estimating the state of a partially observable Markov process from a sequence of observations---is one of the most widely studied problems in control theory, AI, and computational statistics. Exact computation of the posterior distribution is generally intractable for large discrete systems and for nonlinear continuous systems, so a good deal of effort has gone into developing robust approximation algorithms. This paper describes a simple stochastic approximation algorithm for filtering called {em decayed MCMC}. The algorithm applies Markov chain Monte Carlo sampling to the space of state trajectories using a proposal distribution that favours flips of more recent state variables. The formal analysis of the algorithm involves a generalization of standard coupling arguments for MCMC convergence. We prove that for any ergodic underlying Markov process, the convergence time of decayed MCMC with inverse-polynomial decay remains bounded as the length of the observation sequence grows. We show experimentally that decayed MCMC is at least competitive with other approximation algorithms such as particle filtering.

The paper presents an iterative version of join-tree clustering that applies the message passing of join-tree clustering algorithm to join-graphs rather than to join-trees, iteratively. It is inspired by the success of Pearl's belief propagation algorithm as an iterative approximation scheme on one hand, and by a recently introduced mini-clustering i. success as an anytime approximation method, on the other. The proposed Iterative Join-graph Propagation IJGP belongs to the class of generalized belief propagation methods, recently proposed using analogy with algorithms in statistical physics. Empirical evaluation of this approach on a number of problem classes demonstrates that even the most time-efficient variant is almost always superior to IBP and MC i, and is sometimes more accurate by as much as several orders of magnitude.

We present new algorithms for inference in credal networks --- directed acyclic graphs associated with sets of probabilities. Credal networks are here interpreted as encoding strong independence relations among variables. We first present a theory of credal networks based on separately specified sets of probabilities. We also show that inference with polytrees is NP-hard in this setting. We then introduce new techniques that reduce the computational effort demanded by inference, particularly in polytrees, by exploring separability of credal sets.

An increasing number of applications require real-time reasoning under uncertainty with streaming input. The temporal (dynamic) Bayes net formalism provides a powerful representational framework for such applications. However, existing exact inference algorithms for dynamic Bayes nets do not scale to the size of models required for real world applications which often contain hundreds or even thousands of variables for each time slice. In addition, existing algorithms were not developed with real-time processing in mind. We have developed a new computational approach to support real-time exact inference in large temporal Bayes nets. Our approach tackles scalability by recognizing that the complexity of the inference depends on the number of interface nodes between time slices and by exploiting the distinction between static and dynamic nodes in order to reduce the number of interface nodes and to factorize their joint probability distribution. We approach the real-time issue by organizing temporal Bayes nets into static representations, and then using the symbolic probabilistic inference algorithm to derive analytic expressions for the static representations. The parts of these expressions that do not change at each time step are pre-computed. The remaining parts are compiled into efficient procedural code so that the memory and CPU resources required by the inference are small and fixed.