Skat is Germany's national card game played by millions of players around the world. In this paper, we present the world's first computer skat player that plays at the level of human experts. This performance is achieved by improving state evaluations using game data produced by human players and by using these state evaluations to perform inference on the unobserved hands of opposing players. Our results demonstrate the gains from adding inference to an imperfect information game player and show that training on data from average human players can result in expert-level playing strength.

We present an algorithm for selecting an appropriate abstraction when learning a new skill. We show empirically that it can consistently select an appropriate abstraction using very little sample data, and that it significantly improves skill learning performance in a reasonably large real-valued reinforcement learning domain.

We introduce a framework so that communities can exchange reputation information about agents in environments where agents are migrating between communities. We view the acquisition of the reputation information as a purchase and focus on the design of a payment function to facilitate the payment for information in a way that motivates communities to truthfully report reputation information for agents. We prove that in our proposed framework, honesty is the optimal policy and demonstrate the value of using a payment-function approach for the exchange of reputation information about agents between communities in multiagent environments. Using our payment function, each community is strengthened: it is able to reason more effectively about which agents to accept and can enjoy agents that are motivated to contribute strongly to the benefit of the community.

Machine translation services available on the Web are becoming increasingly popular. However, a pivot translation service is required to realize translations between non-English languages by cascading different translation services via English. As a result, the meaning of words often drifts due to the inconsistency , asymmetry and intransitivity of word selections among translation services. In this paper, we propose context-based coordination to maintain the consistency of word meanings during pivot translation services. First, we propose a method to automatically generate multilingual equivalent terms based on bilingual dictionaries and use generated terms to propagate context among combined translation services. Second, we show a multiagent architecture as one way of implementation, wherein a coordinator agent gathers and propagates context from/to a translation agent. We generated trilingual equivalent noun terms and implemented a Japanese-to-German-and-back translation, cascading into four translation services. The evaluation results showed that the generated terms can cover over 58% of all nouns. The translation quality was improved by 40% for all sentences, and the quality rating for all sentences increased by an average of 0.47 points on a five-point scale. These results indicate that we can realize consistent pivot translation services through context-based coordination based on existing services.

Monte Carlo tree search has brought significant improvements to the level of computer players in games such as Go, but so far it has not been used very extensively in games of strongly imperfect information with a dynamic board and an emphasis on risk management and decision making under uncertainty. In this paper we explore its application to the game of Kriegspiel (invisible chess), providing three Monte Carlo methods of increasing strength for playing the game with little specific knowledge. We compare these Monte Carlo agents to the strongest known minimax-based Kriegspiel player, obtaining significantly better results with a considerably simpler logic and less domain-specific knowledge.

The rectangle packing problem consists of finding an enclosing rectangle of smallest area that can contain a given set of rectangles without overlap. Our algorithm picks the x-coordinates of all the rectangles before picking any of the y-coordinates. For the x-coordinates, we present a dynamic variable ordering heuristic and an adaptation of a pruning algorithm used in previous solvers. We then transform the rectangle packing problem into a perfect packing problem that has no empty space, and present inference rules to reduce the instance size. For the y-coordinates we search a space that models empty positions as variables and rectangles as values. Our solver is over 19 times faster than the previous state-of-the-art on the largest problem solved to date, allowing us to extend the known solutions for a consecutive-square packing benchmark from N=27 to N=32.

Previous work by Talbot and Osborne [2007] explored the use of randomized storage mechanisms in language modeling. These structures trade a small amount of error for significant space savings, enabling the use of larger language models on relatively modest hardware. Going beyond space efficient count storage, here we present the Talbot Osborne Morris Bloom (TOMB) Counter, an extended model for performing space efficient counting over streams of finite length. Theoretical and experimental results are given, showing the promise of approximate counting over large vocabularies in the context of limited space.

In this paper, we consider semi-supervised classification on evolutionary data, where the distribution of the data and the underlying concept that we aim to learn change over time due to short-term noises and long-term drifting, making a single aggregated classifier inapplicable for long-term classification. The drift is smooth if we take a localized view over the time dimension, which enables us to impose temporal smoothness assumption for the learning algorithm. We first discuss how to carry out such assumption using temporal regularizers defined in a structural way with respect to the Hilbert space, and then derive the online algorithm that efficiently finds the closed-form solution to the classification functions. Experimental results on real-world evolutionary mailing list data demonstrate that our algorithm outperforms classical semi-supervised learning algorithms in both algorithmic stability and classification accuracy.

The web holds tremendous potential as a source of training data for visual classification. However, web images must be correctly indexed and labeled before this potential can be realized. Accordingly, there has been considerable recent interest in collecting imagery from the web using image search engines to build databases for object and scene recognition research. While search engines can provide rough sets of image data, results are noisy and this leads to problems when training classifiers. In this paper we propose a semi-supervised model for automatically collecting clean example imagery from the web. Our approach includes both visual and textual web data in a unified framework. Minimal supervision is enabled by the selective use of generative and discriminative elements in a probabilistic model and a novel learning algorithm. We show through experiments that our model discovers good training images from the web with minimal manual work. Classifiers trained using our method significantly outperform analogous baseline approaches on the Caltech-256 dataset.

In Model-Based Diagnosis (MBD), the problem of computing a diagnosis in a strong-fault model (SFM) is computationally much harder than in a weak-fault model (WFM). For example, in propositional Horn models, computing the first minimal diagnosis in a weak-fault model (WFM) is in P but is NP-hard for strong-fault models. As a result, SFM problems of practical significance have not been studied in great depth within the MBD community. In this paper we describe an algorithm that renders the problem of computing a diagnosis in several important SFM subclasses no harder than a similar computation in a WFM. We propose an approach for efficiently computing minimal diagnoses for these subclasses of SFM that extends existing conflict-based algorithms like GDE (Sherlock) and CDA*. Experiments on ISCAS85 combinational circuits show (1) inference speedups with CDA* of up to a factor of 8, and (2) an average of 28% reduction in the average conflict size, at the price of an extra low-polynomial-time consistency check for a candidate diagnosis.