The classifiers are based on dyadic classification trees (DCTs), which involve adaptively pruned partitions of the feature space. A key aspect of DCTs is their spatial adaptivity, which enables local (ratherthan global) fitting of the decision boundary. Our risk analysis involves a spatial decomposition of the usual concentration inequalities, leading to a spatially adaptive, data-dependent pruning criterion. For any distribution on (X, Y) whose Bayes decision boundary behaves locally like a Lipschitz smooth function, we show that the DCT error converges to the Bayes error at a rate within a logarithmic factor of the minimax optimal rate.

We propose an approach to learning the semantics of images which allows usto automatically annotate an image with keywords and to retrieve images based on text queries. We do this using a formalism that models the generation of annotated images. We assume that every image is divided intoregions, each described by a continuous-valued feature vector. Given a training set of images with annotations, we compute a joint probabilistic modelof image features and words which allow us to predict the probability of generating a word given the image regions. This may be used to automatically annotate and retrieve images given a word as a query. Experiments show that our model significantly outperforms the best of the previously reported results on the tasks of automatic image annotation and retrieval.

We describe a new approximation algorithm for solving partially observable MDPs. Our bounded policy iteration approach searches through the space of bounded-size, stochastic finite state controllers, combining several advantages of gradient ascent (efficiency, search through restricted controller space) and policy iteration (less vulnerability to local optima).

As part of an environmental observation and forecasting system, sensors deployed in the Columbia RIver Estuary (CORIE) gather information on physical dynamics and changes in estuary habitat. Ofthese, salinity sensors are particularly susceptible to biofouling, whichgradually degrades sensor response and corrupts critical data. Automatic fault detectors have the capability to identify bio-fouling early and minimize data loss. Complicating the development ofdiscriminatory classifiers is the scarcity of bio-fouling onset examples and the variability of the bio-fouling signature. To solve these problems, we take a novelty detection approach that incorporates a parameterized bio-fouling model. These detectors identify the occurrence of bio-fouling, and its onset time as reliably as human experts. Real-time detectors installed during the summer of2001 produced no false alarms, yet detected all episodes of sensor degradation before the field staff scheduled these sensors for cleaning. From this initial deployment through February 2003, our bio-fouling detectors have essentially doubled the amount of useful data coming from the CORIE sensors.

Joseph then to students who are actively February 19: IS Demos proposals E. Beck (joseph.beck@cmu.edu)

Satellite domains are becoming a fashionable area of research within the AI community due to the complexity of the problems that satellite domains need to solve. With the current U.S. and European focus on launching satellites for communication, broadcasting, or localization tasks, among others, the automatic control of these machines becomes an important problem. Many new techniques in both the planning and scheduling fields have been applied successfully, but still much work is left to be done for reliable autonomous architectures. The purpose of this article is to present CONSAT, a real application that plans and schedules the performance of nominal operations in four satellites during the course of a year for a commercial Spanish satellite company, HISPASAT. For this task, we have used an AI domain-independent planner that solves the planning and scheduling problems in the HISPASAT domain thanks to its capability of representing and handling continuous variables, coding functions to obtain the operators' variable values, and the use of control rules to prune the search. We also abstract the approach in order to generalize it to other domains that need an integrated approach to planning and scheduling.

(Niles and Pease 2001). Considering that tremendous scheduling that are robust in the face of realworld progress has been made in commonsense reasoning concerns like time zones, daylight savings in specialized topics such as thermodynamics time, and international calendar variations. in physical systems (Collins and Forbus 1989), it is surprising that our best content theories Given the importance of an ontology of of people are still struggling to get past time across so many different commonsense simple notions of belief and intentionality (van der Hoek and Wooldridge 2003). However, search is the generation of competency theories systems that can successfully reason about that have a degree of depth necessary to solve people are likely to be substantially more valuable inferential problems that people are easily able than those that reason about thermodynamics to handle. in most future applications. Yet competency in content theories is only Content theories for reasoning about people half of the challenge. Commonsense reasoning are best characterized collectively as a theory of in AI theories will require that computers not commonsense psychology, in contrast to those only make deep humanlike inferences but also that are associated with commonsense (naïve) ensure that the scope of these inferences is as physics. The scope of commonsense physics, broad as humans can handle, as well. That is, best outlined in Patrick Hayes's first and second in addition to competency, content theories will "Naïve Physics Manifestos" (Hayes 1979, need adequate coverage over the full breadth of 1984), includes content theories of time, space, concepts that are manipulated in human-level physical entities, and their dynamics. It is only by achieving psychology, in contrast, concerns all some adequate level of coverage that we of the aspects of the way that people think they can begin to construct reasoning systems that think. It should include notions of plans and integrate fully into real-world AI applications, goals, opportunities and threats, decisions and where pragmatic considerations and expressive preferences, emotions and memories, along user interfaces raise the bar significantly.

The Fourteenth International Conference on Automated Planning and Scheduling (ICAPS-04) was held in Canada in June of 2004. It covered the latest theoretical and empirical advances in planning and scheduling. The conference program consisted of tutorials, workshops, a doctoral consortium, and three days of technical paper presentations in a single plenary track, one day of which was jointly organized with the Ninth International Conference on Principles of Knowledge Representation and Reasoning. ICAPS-04 also hosted the International Planning Competition, including a classical track and a newly formed probabilistic track. This report describes the conference in more detail.

AAAI's Nineteenth National Conference on Artificial Intelligence (AAAI-04) filled the top floor of the San Jose Convention Center from July 25-29, 2004. The week's program was full of recent advances in many different AI research areas, as well as emerging applications for AI. Within the various topics discussed at the conference, a number of strategic domains emerged where AI is being harnessed, including counterterrorism, space exploration, robotics, the Web, health care, scientific research, education, and manufacturing.

We introduce a stochastic graph-based method for computing relative importance of textual units for Natural Language Processing. We test the technique on the problem of Text Summarization (TS). Extractive TS relies on the concept of sentence salience to identify the most important sentences in a document or set of documents. Salience is typically defined in terms of the presence of particular important words or in terms of similarity to a centroid pseudo-sentence. We consider a new approach, LexRank, for computing sentence importance based on the concept of eigenvector centrality in a graph representation of sentences. In this model, a connectivity matrix based on intra-sentence cosine similarity is used as the adjacency matrix of the graph representation of sentences. Our system, based on LexRank ranked in first place in more than one task in the recent DUC 2004 evaluation. In this paper we present a detailed analysis of our approach and apply it to a larger data set including data from earlier DUC evaluations. We discuss several methods to compute centrality using the similarity graph. The results show that degree-based methods (including LexRank) outperform both centroid-based methods and other systems participating in DUC in most of the cases. Furthermore, the LexRank with threshold method outperforms the other degree-based techniques including continuous LexRank. We also show that our approach is quite insensitive to the noise in the data that may result from an imperfect topical clustering of documents.