The leading approach for computing strong game-theoretic strategies in large imperfect-information games is to first solve an abstracted version of the game offline, then perform a table lookup during game play. We consider a modification to this approach where we solve the portion of the game that we have actually reached in real time to a greater degree of accuracy than in the initial computation. We call this approach endgame solving. Theoretically, we show that endgame solving can produce highly exploitable strategies in some games; however, we show that it can guarantee a low exploitability in certain games where the opponent is given sufficient exploitative power within the endgame. Furthermore, despite the lack of a general worst-case guarantee, we describe many benefits of endgame solving. We present an efficient algorithm for performing endgame solving in large imperfect-information games, and present a new variance-reduction technique for evaluating the performance of an agent that uses endgame solving. Experiments on no-limit Texas Hold'em show that our algorithm leads to significantly stronger performance against the strongest agents from the 2013 AAAI Annual Computer Poker Competition.

An important aspect of health science is communicating research findings to the public. The media is a critical instrument in disseminating research. Yet the process by which a scientific article becomes “newsworthy” is not well understood. In this study, we use large-scale text analysis to characterize the content features of articles that are predictive of newsworthiness. We experiment with two novel corpora: (i) 28,910 articles from a di- verse range of biomedical and health journals, of which 1,343 were covered by the news agency Reuters, and (ii) 10,760 articles from the JAMA journals, of which 846 were given press releases by the journal editors. We show that media coverage can be predicted reasonably well: logistic regression achieves mean AUCs of 0.783 and 0.882 on the Reuters and JAMA datasets, respec- tively. We present and discuss interesting findings con- cerning the most predictive content features.

How can knowing the location of my friends be used to more accurately predict my location? This paper explores socially-aware location prediction under a particularly challenging setting where the underlying interactions and social network are unknown and must be inferred over continuous spatiotemporal data. Our method samples inferred network topology using a linear regression model to predict future individual locations. We present an in-depth empirical study comparing different network models and network sampling regimes under a bootstrapped sampling baseline. Furthermore, our qualitative analysis demonstrates the value of social information in population mobility modeling under our application’s challenges.

In this work we present a theoretical approach (not currently implemented), to the problem of predicting agent behavior. The ultimate goal of this work is to learn models that can be used to predict the future actions of intelligent agents, based on previously recorded data on those agents’ behavior. We believe that we can improve the predictive accuracy of our models by assuming that an agent reasons about the actions it takes, and trying to explicitly model that reasoning process. Here, we model an agent’s reasoning process as a form of Monte-Carlo search, and attempt to learn a state evaluation function that, when used with this planning algorithm, yields a similar distribution of actions given the current state of the world as we observe in the data. While it is simple to simulate Monte-Carlo search given an evaluation function, it is much more difficult to determine an evaluation function that will generate a certain behavior. Here we will use Expectation-Maximization to find a maximum likelihood estimate of the parameters of the evaluation function, treating the actual steps taken in planning each action as unobserved data.

While most work on trajectory mining is applied to pre- dict movements of mobile users, in this paper we consider a more general problem of building behavior models of users from their state-action trajectories. We assume that the user behavior can be compactly modeled as a Probabilistic State-Dependent Grammar (PSDG) which represents a hierarchical policy. The key problem is that while the states and actions of the user are directly observed, his intentional structure is not. We propose to learn the user’s policy from a set of selected trajectories and intention queries at selected states in the trajectory. Our main contributions are an algorithm for learning hierarchical policies from state-action trajectories, and principled heuristics for selecting suitable trajectories and intention queries. Experiments in multiple domains show that our approach is effective and more sample-efficient than learning non-hierarchical policies.

Systems biology, the study of the intricate, ramified, com-plex and interacting mechanisms underlying life, often proves too complex for unaided human understanding, even by groups of people working together. This difficulty is ex-acerbated by the high volume of publications in molecular biology. The Big C (‘C’ for Cyc) is a system designed to (semi-)automatically acquire, integrate, and use complex mechanism models, specifically related to cancer biology, via automated reading and a hyper-detailed refinement pro-cess resting on Cyc’s logical representations and powerful inference mechanisms. We aim to assist cancer research and treatment by achieving elements of biologist-level reason-ing, but with the scale and attention to detail that only com-puter implementations can provide.

We introduce the novel task of identifying important citations in scholarly literature, i.e., citations that indicate that the cited work is used or extended in the new effort. We believe this task is a crucial component in algorithms that detect and follow research topics and in methods that measure the quality of publications.We model this task as a supervised classification problem at two levels of detail: a coarse one with classes (important vs. non-important), and a more detailed one with four importance classes. We annotate a dataset of approximately 450 citations with this information, and release it publicly. We propose a supervised classification approach that addresses this task with a battery of features that range from citation counts to where the citation appears in the body of the paper, and show that,our approach achieves a precision of 65% for a recall of 90%.

Recommender systems provide personalized item suggestions by identifying patterns in past user-item preferences. Most existing approaches for recommender systems work on a single domain, i.e., use user preferences from one domain and recommend items from the same domain. Recently, some recommendation models have been proposed to use user preferences from multiple related item source domains to improve recommendation accuracy for a target item domain, an area of research known as cross-domain recommender systems. One typical assumption in these systems is that users, items, and user preferences for items are similar across domains. In this paper, we introduce a new cross-domain recommendation problem which does not meet this typical assumption. For example, for some scientometric datasets, when the objective is to recommend co-authors, conferences, and references, respectively, to authors, although the users are similar across domains, the items and user-item preferences are different. To address this problem, we propose two approaches to aggregate knowledge from multiple domains. Our approaches allow us to control the knowledge transferred between domains. Experimental results on a DBLP subset show that the proposed cross-domain approaches are helpful in improving recommendation accuracy as compared to single domain approaches.

Among the languages used for representing goals, actions and their consequences on the world for decision making and planning, GDL (Game Description Language) has the ability to represent complex actions in potentially uncertain and competitive environments. The aim of this paper is to exploit stochastic constraint networks in order to provide compact representations of strategic games, and to identify optimal policies in those games with generic forward checking method. From this perspective, we develop a compiler allowing to translate games, described in GDL, into instances of the Stochastic Constraint Optimization Problem (SCSP). Our compiler is proved correct for the class GDL of games with complete information and oblivious environment. The interest of our approach is illustrated by solving several GDL games with a SCSP solver.

The paper explores the potential of weighted best-first search schemes as anytime optimization algorithms for solving graphical models tasks such as MPE (Most Probable Explanation) or MAP (Maximum a Posteriori) and WCSP (Weighted Constraint Satisfaction Problem). While such schemes were widely investigated for path-finding tasks, their application for graphical models was largely ignored, possibly due to their memory requirements. Compared to the depth-first branch and bound, which has long been the algorithm of choice for optimization in graphical models, a valuable virtue of weighted best-first search is that they are w-optimal, i.e. when terminated, they return a solution cost C and a weight w, such that C < = wC*, where C* is the optimal cost. We report on a significant empirical evaluation, demonstrating the usefulness of weighted best-first search as approximation anytime schemes (that have suboptimality bounds) and compare against one of the best depth-first branch and bound solver to date. We also investigate the impact of different heuristic functions on the behaviour of the algorithms.