CSP solvers usually search a partial assignment search tree.We present a new formalization for CSP solvers, which spansa conceptually different search tree, where each node representssubsets of the original domains for the variables. We experimentwith a simple backtracking algorithm for this searchtree and show that it outperforms a simple backtracking algorithmon the traditional search tree in many cases.

We propose a new perspective on the automated design of combinatorial search algorithms through an approach that operates at a much higher semantic level than previous algorithm configurators do. Instead of blindly tuning numerical or categorical parameters based on black-box optimization or resorting to a handful of predefined strategies, we propose to automatically search over compositions of search strategies using a light-weight language, while exploiting the semantic knowledge of the modeling language itself to guide the configuration process. Although somewhat reminiscent of the old AI vision that machines will be able to program themselves to solve novel tasks, we believe that the idea restricted to this simple but powerful search language has a chance of success in practice.

Photo ranking algorithms aim to quantify features within photographs to determine aesthetic quality and learn user preferences over these features. However, current benchmark corpora of photographs contain non-quantifiable contextual information. Moreover, they do not control the variance of quantifiable features. The recently released Panorama data set contains annotated synthetic images with controlled contextual features. The images lay along the full range of quantifiable features. This paper focuses on improving the performance of a predictive learning model trained on the pairwise preferences collected on these images. Predictive models were trained on individual as well as group preferences. Feature selection improved prediction accuracy from the 67% earlier reported to 91% accuracy on the group preference. On average, individual raters' preferences were predicted with 87% accuracy. Top features varied widely among individuals. However, the most frequent top three features were tilted horizon line, cropped objects, and number of objects.

An important problem in AI is to construct high-quality heuristics for optimal search. Recently, the Euclidean heuristic (EH) has been proposed, which embeds a state space graph into a Euclidean space and uses Euclidean distances as approximations for the graph distances. The embedding process leverages recent research results from manifold learning, a subfield in machine learning, and guarantees that the heuristic is provably admissible and consistent. EH has shown good performance and memory efficiency in comparison to other existing heuristics. Our recent works have further improved the scalability and quality of EH. In this short paper, we present our latest progress on applying EH to problems in planning formalisms, which provide richer semantics than the simple state-space graph model. In particular, we improve EH by exploiting the landmark structure derived from the SAS+ planning formalism.

Modularity is essential for modeling large-scale practical applications. We propose modular logic programs as a modular version of answer set programming and study the relationship of our formalism to an earlier concept of lp-modules.

We build turn-taking models for autonomous characters in language-based interactions with small groups of children. Two models explore the use of support vector machines given the same multimodal features, but different methods for collecting turn-taking labels.

The effectiveness of supervised feature selection degrades in low training data scenarios. We propose to alleviate this problem by augmenting per-task feature selection with joint feature selection over multiple tasks. Our algorithm builds on the assumption that different tasks have shared structure which could be utilized to cope with data sparsity. The proposed trace-ratio based model not only selects discriminative features for each task, but also finds features which are discriminative over all tasks. Extensive experiment on different data sets demonstrates the effectiveness of our algorithm in low training data scenarios.

Distributed Constraint Problem solving represents a fundamental research area in distributed artificial intelligence and multi-agent systems. The constraint density, or the ratio of the number of constraints to the number of variables, determines the difficulty of either finding a solution or minimizing the set of variable assignment conflicts. Reducing density typically reduces difficulty. We present a fully distributed technique for reducing the effective density of constraint graphs, called Virtual Structure Reduction (VSR). The VSR technique leverages the occurrence of variables that must be assigned the same value based on shared constraints and can improve solver performance using existing algorithms. We discuss our Distributed Constraint Optimization Problem (DCOP) solver, integrated with the Distributed Stochastic Algorithm (DSA), called VSR-DSA. The VSR-DSA algorithm demonstrates performance gains vs DSA in both solution quality and time on 3-coloring problems.

In this paper, we address the issue of bilingual sentiment lexicon learning(BSLL) which aims to automatically and simultaneously generate sentiment words for two languages. The underlying motivation is that sentiment information from two languages can perform iterative mutual-teaching in the learning procedure. We propose to develop two classifiers to determine the sentiment polarities of words under a co-training framework, which makes full use of the two-view sentiment information from the two languages. The word alignment derived from the parallel corpus is leveraged to design effective features and to bridge the learning of the two classifiers. The experimental results on English and Chinese languages show the effectiveness of our approach in BSLL.

The Web is currently shifting from data on linked Web pages towards less interlinked data in social networks on the Web. Therefore, rather than being based on the link structure between Web pages, the ranking of search results needs to be based on something new. We believe that it can be based on user preferences and ontological background knowledge, as a means to personalized access to information. There are many approaches to preference representation and reasoning in the literature. The most prominent qualitative ones are perhaps CP-nets. Their clear graphical structure unifies an easy representation of preferences with nice properties when computing the best outcome. In this paper, we introduce ontological CP-nets, where the knowledge domain has an ontological structure, i.e., the values of the variables are constrained relative to an underlying ontology. We show how the computation of Pareto optimal outcomes for such ontological CP-nets can be reduced to the solution of constraint satisfaction problems. We also provide several complexity and tractability results.