Exchangeability is a central notion in statistics and probability theory. The assumption that an infinite sequence of data points is exchangeable is at the core of Bayesian statistics. However, finite exchangeability as a statistical property that renders probabilistic inference tractable is less well-understood. We develop a theory of finite exchangeability and its relation to tractable probabilistic inference. The theory is complementary to that of independence and conditional independence. We show that tractable inference in probabilistic models with high treewidth and millions of variables can be explained with the notion of finite (partial) exchangeability. We also show that existing lifted inference algorithms implicitly utilize a combination of conditional independence and partial exchangeability.

There are various algorithms for finding a Bayesian networkstructure (BNS) that is optimal with respect to a given scoring function. No single algorithm dominates the others in speed, and, given a problem instance, it is a priori unclear which algorithm will perform best and how fast it will solve the problem. Estimating the runtimes directly is extremely difficult as they are complicated functions of the instance. The main contribution of this paper is characterization of the empirical hardness of an instance for a given algorithm based on a novel collection of non-trivial, yet efficiently computable features. Our empirical results, based on the largest evaluation of state-of-the-art BNS learning algorithms to date, demonstrate that we can predict the runtimes to a reasonable degree of accuracy, and effectively select algorithms that perform well on a particular instance. Moreover, we also show how the results can be utilized in building a portfolio algorithm that combines several individual algorithms in an almost optimal manner.

One-class classification approaches have been proposed in the literature to learn classifiers from examples of only one class. But these approaches are not directly applicable to relational domains due to their reliance on a feature vector or a distance measure. We propose a non-parametric relational one-class classification approach based on first-order trees. We learn a tree-based distance measure that iteratively introduces new relational features to differentiate relational examples. We update the distance measure so as to maximize the one-class classification performance of our model. We also relate our model definition to existing work on probabilistic combination functions and density estimation. We experimentally show that our approach can discover relevant features and outperform three baseline approaches.

Monte Carlo tree search (MCTS) algorithms are a popular approach to online decision-making in Markov decision processes (MDPs). These algorithms can, however, perform poorly in MDPs with high stochastic branching factors. In this paper, we study state aggregation as a way of reducing stochastic branching in tree search. Prior work has studied formal properties of MDP state aggregation in the context of dynamic programming and reinforcement learning, but little attention has been paid to state aggregation in MCTS. Our main result is a performance loss bound for a class of value function-based state aggregation criteria in expectimax search trees. We also consider how to construct MCTS algorithms that operate in the abstract state space but require a simulator of the ground dynamics only. We find that trajectory sampling algorithms like UCT can be adapted easily, but that sparse sampling algorithms present difficulties. As a proof of concept, we experimentally confirm that state aggregation can improve the finite-sample performance of UCT.

A recent breadth-first branch and bound algorithm (BFBnB)for learning Bayesian network structures (Maloneet al. 2011) uses two bounds to prune the searchspace for better efficiency; one is a lower bound calculatedfrom pattern database heuristics, and the otheris an upper bound obtained by a hill climbing search.Whenever the lower bound of a search path exceeds theupper bound, the path is guaranteed to lead to suboptimalsolutions and is discarded immediately. This paperintroduces methods for tightening the bounds. Thelower bound is tightened by using more informed variablegroupings when creating the pattern databases, andthe upper bound is tightened using an anytime learningalgorithm. Empirical results show that these boundsimprove the efficiency of Bayesian network learning bytwo to three orders of magnitude.

In this paper we develop an algorithm to find the k-best equivalence classes of Bayesian networks. Our algorithm is capable of finding much more best DAGs than the previous algorithm that directly finds the k-best DAGs (Tian, He and Ram 2010). We demonstrate our algorithm in the task of Bayesian model averaging. Empirical results show that our algorithm significantly outperforms the k-best DAG algorithm in both time and space to achieve the same quality of approximation. Our algorithm goes beyond the maximum-a-posteriori (MAP) model by listing the most likely network structures and their relative likelihood and therefore has important applications in causal structure discovery.

In causal inference, all methods of model learning rely on testable implications, namely, properties of the joint distribution that are dictated by the model structure. These constraints, if not satisfied in the data, allow us to reject or modify the model. Most common methods of testing a linear structural equation model (SEM) rely on the likelihood ratio or chi-square test which simultaneously tests all of the restrictions implied by the model. Local constraints, on the other hand, offer increased power (Bollen and Pearl, 2013; McDonald, 2002) and, in the case of failure, provide the modeler with insight for revising the model specification. One strategy of uncovering local constraints in linear SEMs is to search for overidentified path coefficients. While these overidentifying constraints are well known, no method has been given for systematically discovering them. In this paper, we extend the half-trek criterion of (Foygel et al., 2012) to identify a larger set of structural coefficients and use it to systematically discover overidentifying constraints. Still open is the question of whether our algorithm is complete.

Fitted Q-iteration (FQI) stands out among reinforcement learning algorithms for its flexibility and ease of use. FQI can be combined with any regression method, and this choice determines the algorithm's statistical and computational properties. The combination of FQI with an ensemble of regression trees gives rise to an algorithm, FQIT, that is computationally efficient, scalable to high dimensional spaces, and robust to noise. Despite its nice properties and good performance in practice, FQIT also has some limitations: the fact that an ensemble of trees must be constructed (or updated) at each iteration confines the algorithm to the batch scenario. This paper aims to address this specific issue. Based on a strategy recently proposed in the literature, called the stochastic-factorization trick, we propose a modification of FQIT that makes it fully incremental, and thus suitable for on-line learning. We call the resulting method tree-based stochastic factorization (TBSF). We derive upper bounds for the difference between the value functions computed by FQIT and TBSF, and also show in which circumstances the approximations coincide. A series of computational experiments is presented to illustrate the properties of TBSF and to show its usefulness in practice, including a medical problem involving the treatment of patients infected with HIV.

Selection bias is caused by preferential exclusion of units from the samples and represents a major obstacle to valid causal and statistical inferences; it cannot be removed by randomized experiments and can rarely be detected in either experimental or observational studies. In this paper, we provide complete graphical and algorithmic conditions for recovering conditional probabilities from selection biased data. We also provide graphical conditions for recoverability when unbiased data is available over a subset of the variables. Finally, we provide a graphical condition that generalizes the backdoor criterion and serves to recover causal effects when the data is collected under preferential selection.

Inference in large scale graphical models is an important task in many domains, and in particular probabilistic relational models (e.g. Markov logic networks). Such models often exhibit considerable symmetry, and it is a challenge to devise algorithms that exploit this symmetry to speed up inference. Recently, the automorphism group has been proposed to formalize mathematically what "exploiting symmetry" means. However, obtaining symmetry derived from automorphism is GI-hard, and consequently only a small fraction of the symmetry is easily available for effective employment. In this paper, we improve upon efficiency in two ways. First, we introduce the Cluster Signature Graph (CSG), a platform on which greater portions of the symmetries can be revealed and exploited. CSGs classify clusters of variables by projecting relations between cluster members onto a graph, allowing for the efficient pruning of symmetrical clusters even before their generation. Second, we introduce a novel framework based on CSGs for the Sherali-Adams hierarchy of linear program (LP) relaxations, dedicated to exploiting this symmetry for the benefit of tight Maximum A Posteriori (MAP) approximations. Combined with the pruning power of CSG, the framework quickly generates compact formulations for otherwise intractable LPs, as demonstrated by several empirical results.