Advances in Artificial Intelligence

This paper describes a reactive navigation method for autonomous agents such as robots or actors in virtual worlds, based on novel

dynamic tangent

obstacle representations, resulting in exceptionally successful, geometrically sensitive navigation. The method employs three levels of abstraction, treating each obstacle entity as an

obstacle-valued function

; this treatment enables extraordinary flexibility without pre-computation or deliberation, applying to all obstacles regardless of shape, including non-convex, polygonal, or arc-shaped obstacles in dynamic environments. The unconventional levels of abstraction and the geometric details of dynamic tangent representations are the primary contributions of this work, supporting smooth navigation even in scenarios with curved shapes, such as circular and figure-eight shaped tracks, or in environments requiring complex, winding paths.

Given a finite domain, grounding is the the process of creating a variable-free first-order formula equivalent to a first-order sentence. As the first-order sentences can be used to describe a combinatorial search problem, efficient grounding algorithms would help in solving such problems effectively and makes advanced solver technology (such as SAT) accessible to a wider variety of users. One promising method for grounding is based on the relational algebra from the field of Database research. In this paper, we describe the extension of this method to ground formulas of first-order logic extended with arithmetic, expansion functions and aggregate operators. Our method allows choice of particular CNF representations for complex constraints, easily.

This paper introduces the concept of

Moving Object

(MO) modelling as a means of managing the uncertainty in the location tracking of human moving objects travelling on a network. For previous movements of the MOs, the uncertainty stems from the discrete nature of location tracking systems, where gaps are created among the location reports. Future locations of MOs are, by definition, uncertain. The objective is to maximize the estimation accuracy while minimizing the operating costs.

In this paper, we investigate an unsupervised approach to Relation Extraction to be applied in the context of automatic generation of multiple-choice questions (MCQs). MCQs are a popular large-scale assessment tool making it much easier for test-takers to take tests and for examiners to interpret their results. Our approach to the problem aims to identify the most important semantic relations in a document without assigning explicit labels to them in order to ensure broad coverage, unrestricted to predefined types of relations. In this paper, we present an approach to learn semantic relations between named entities by employing a dependency tree model. Our findings indicate that the presented approach is capable of achieving high precision rates, which are much more important than recall in automatic generation of MCQs, and its enhancement with linguistic knowledge helps to produce significantly better patterns. The intended application for the method is an e-Learning system for automatic assessment of students’ comprehension of training texts; however it can also be applied to other NLP scenarios, where it is necessary to recognise the most important semantic relations without any prior knowledge as to their types.

We introduce Satisfiable Random High Degree Subgraph Isomorphism Generator(SRHD-SGI), a variation of the Satisfiable Random Subgraph Isomorphism Generator (SR-SGI). We use the direct encoding to translate the SRHD-SGI instances into Satisfiable SAT instances. We present empirical evidence that the new model preserves the main characteristics of SAT encoded SR-SGI: easy-hard-easy pattern of evolution and exponential growth of empirical hardness. Our experiments indicate that SAT encoded SRHD-SGI instances are empirically harder than their SR-SGI counterparts. Therefore we conclude that SRHD-SGI is an improved generator of satisfiable SAT instances.

Existing preference prediction techniques can require that an entire preference structure be constructed for a user. These structures, such as Conditional Outcome Preference Networks (COP-nets), can grow exponentially in the number of attributes describing the outcomes. In this paper, a new approach for constructing COP-nets, using A* search, is introduced. Using this approach, partial COP-nets can be constructed on demand instead of generating the entire structure. Experimental results show that the new method yields enormous savings in time and memory requirements, with only a modest reduction in prediction accuracy.

PROAFTN belongs to Multiple-Criteria Decision Aid (MCDA) paradigm and requires a several set of parameters for the purpose of classification. This study proposes a new inductive approach for obtaining these parameters from data. To evaluate the performance of developed learning approach, a comparative study between PROAFTN and a decision tree in terms of their learning methodology, classification accuracy, and interpretability is investigated in this paper. The major distinguished property of Decision tree is that its ability to generate classification models that can be easily explained. The PROAFTN method has also this capability, therefore avoiding a black box situation. Furthermore, according to the proposed learning approach in this study, the experimental results show that PROAFTN strongly competes with ID3 and C4.5 in terms of classification accuracy.

In this paper, we adopt a supervised machine learning approach to recognize six basic emotions (anger, disgust, fear, happiness, sadness and surprise) using a heterogeneous emotion-annotated dataset which combines news headlines, fairy tales and blogs. For this purpose, different features sets, such as bags of words, and N-grams, were used. The Support Vector Machines classifier (SVM) performed significantly better than other classifiers, and it generalized well on unseen examples.

In this paper, we propose the use of semantic information for the task of answering complex questions. We use the Extended String Subsequence Kernel (ESSK) to perform similarity measures between sentences in a graph-based random walk framework where semantic information is incorporated by exploiting the word senses. Experimental results on the DUC benchmark datasets prove the effectiveness of our approach.

This paper describes a method to perform automated semantic annotation of named entities contained in large corpora. The semantic annotation is made in the context of the Semantic Web. The method is based on an algorithm that compares the set of words that appear before and after the name entity with the content of Wikipedia articles, and identifies the more relevant one by means of a similarity measure. It then uses the link that exists between the selected Wikipedia entry and the corresponding RDF description in the Linked Data project to establish a connection between the named entity and some URI in the Semantic Web. We present our system, discuss its architecture, and describe an algorithm dedicated to ontological disambiguation of named entities contained in large-scale corpora. We evaluate the algorithm, and present our results.

In this paper, we learn the components of dialogue POMDP models from data. In particular, we learn the states, observations, as well as transition and observation functions based on a Bayesian latent topic model using unannotated human-human dialogues. As a matter of fact, we use the Bayesian latent topic model in order to learn the intentions behind user’s utterances. Similar to recent dialogue POMDPs, we use the discovered user’s intentions as the states of dialogue POMDPs. However, as opposed to previous works, instead of using some keywords as POMDP observations, we use some meta observations based on the learned user’s intentions. As the number of meta observations is much less than the actual observations, i.e. the number of words in the dialogue set, the POMDP learning and planning becomes tractable. The experimental results on real dialogues show that the quality of the learned models increases by increasing the number of dialogues as training data. Moreover, the experiments based on simulation show that the introduced method is robust to the ASR noise level.

The field of Reinforcement Learning (RL) in machine learning relates significantly to the domains of classical and instrumental conditioning in psychology, which give an understanding of biology’s approach to RL. In recent years, there has been a thrust to correlate some machine learning RL algorithms with brain structure and function, a benefit to both fields. Our focus has been on one such structure, the striatum, from which we have built a general model. In machine learning terms, this model is equivalent to a value-function approximator (VFA) that learns according to Temporal Difference error. In keeping with a biological approach to RL, the present work seeks to evaluate the robustness of this striatum-based VFA using biological criteria. We selected five classical conditioning tests to expose the learning accuracy and efficiency of the VFA for simple state-value associations. Manually setting the VFA’s many parameters to reasonable values, we characterize it by varying each parameter independently and repeatedly running the tests. The results show that this VFA is both capable of performing the selected tests and is quite robust to changes in parameters. Test results also reveal aspects of how this VFA encodes reward value.

This paper explores Answer Set Programming (ASP) for stream reasoning with data retrieved continuously from sensors. We describe a proof-of-concept with an example of using declarative models to recognize car on-road situations.

We consider the problem of strategic decision-making for inshore sailboat racing. This sequential decision-making problem is complicated by the yacht’s dynamics which prevent it from sailing directly into the wind but allow it to sail close to the wind following a zigzag trajectory towards an upwind race marker. A skipper is faced with the problem of sailing the most direct route to this marker whilst minimizing the number of steering manoeuvres that slow down the boat. In this paper, we present a Decision Theoretic model for this decision-making process assuming a fully observable environment and uncertain boat dynamics. We develop a numerical Velocity Prediction Program (VPP) which allows us to predict the yacht’s speed and direction of sail given the wind’s strength and direction as well as the yacht’s angle of attack with respect to the wind. We specify and solve a Markov Decision Process (MPD) using our VPP to estimate the rewards and transition probabilities. We also present a method for modelling the wind flow around landmasses allowing for the computation of strategies in realistic situations. Finally, we evaluate our approach in simulation showing that we can estimate optimal routes for different kinds of yachts and crew performance.

In this work, we present a method for classifying the quality of blog comments using Linear-Chain Conditional Random Fields (CRFs). This approach is found to yield high accuracy on binary classification of high-quality comments, with conversational features contributing strongly to the accuracy. We also present a new corpus of blog data in conversational form, complete with user-generated quality moderation labels from the science and technology news blog Slashdot.

Machine lifelong learning (ML3) is concerned with machines capable of learning and retaining knowledge over time, and exploiting this knowledge to assist new learning. An ML3 system must accurately retain knowledge of prior tasks while consolidating in knowledge of new tasks, overcoming the stability-plasticity problem. A system is presented using a context-sensitive multiple task learning (

cs

MTL) neural network.

cs

MTL uses a single output and additional context inputs for associating examples with tasks. A

cs

MTL-based ML3 system is analyzed empirically using synthetic and real domains. The experiments focus on the effective retention and consolidation of task knowledge using both functional and representational transfer. The results indicate that combining the two methods of transfer serves best to retain prior knowledge, but at the cost of less effective new task consolidation.

This paper describes research methodologies and experimental settings for the task of relation identification and classification between pairs of medical entities, using clinical data. The models that we use represent a combination of lexical and syntactic features, medical semantic information, terms extracted from a vector-space model created using a random projection algorithm, and additional contextual information extracted at sentence-level. The best results are obtained using an SVM classification algorithm with a combination of the above mentioned features, plus a set of additional features that capture the distributional semantic correlation between the concepts and each relation of interest.

This work examines a novel method of developing features to use for machine learning of sentiment analysis and related tasks. This task is frequently approached using a “Bag of Words” representation – one feature for each word encountered in the training data – which can easily involve thousands of features. This paper describes a set of compact features developed by learning scores for words, dividing the range of possible scores into a number of bins, and then generating features based on the distribution of scored words in the document over the bins. This allows for effective learning of sentiment and related tasks with 25 features; in fact, performance was very often slightly better with these features than with a simple bag of words baseline. This vast reduction in the number of features reduces training time considerably on large datasets, and allows for using much larger datasets than previously attempted with bag of words approaches, improving performance.

Semi-supervised learning methods utilize abundant unlabeled data to help to learn a better classifier when the number of labeled instances is very small. A common method is to select and label unlabeled instances that the current classifier has high classification confidence to enlarge the labeled training set and then to update the classifier, which is widely used in two paradigms of semi-supervised learning: self-training and co-training. However, the original labeled instances are more reliable than the self-labeled instances that are labeled by the classifier. If unlabeled instances are assigned wrong labels and then used to update the classifier, classification accuracy will be jeopardized. In this paper, we present a new instance selection method based on the original labeled data (ISBOLD). ISBOLD considers not only the prediction confidence of the current classifier on unlabeled data but also its performance on the original labeled data only. In each iteration, ISBOLD uses the change of accuracy of the newly learned classifier on the original labeled data as a criterion to decide whether the selected most confident unlabeled instances will be accepted to the next iteration or not. We conducted experiments in self-training and co-training scenarios when using Naive Bayes as the base classifier. Experimental results on 26 UCI datasets show that, ISBOLD can significantly improve accuracy and AUC of self-training and co-training.

The Seismic Research Centre, University of the West Indies operates a seismic network that performs suboptimally in detecting, locating, and correctly determining the magnitude of earthquakes due to a diverse constituion of seismometers and the utilization of a site selection process that approximates an educated guess. My work seeks to apply Self-Organizing Maps (SOM) to arrive at the optimal network configuration and aid in site selection.

Advances in high frequency sonar have provided increasing resolution of sea bottom objects, providing higher fidelity sonar data for automated target recognition tools. Here we investigate if advanced techniques in the field of visual object recognition and machine learning can be applied to classify mine-like objects from such sonar data. In particular, we investigate if the recently popular Scale-Invariant Feature Transform (SIFT) can be applied for such high-resolution sonar data. We also follow up our previous approach in applying the unsupervised learning of deep belief networks, and advance our methods by applying a convolutional Restricted Boltzmann Machine (cRBM). Finally, we now use Support Vector Machine (SVM) classifiers on these learned features for final classification. We find that the cRBM-SVM combination slightly outperformed the SIFT features and yielded encouraging performance in comparison to state-of-the-art, highly engineered template matching methods.

An increasing number of data mining domains consider data that can be represented as permutations. Therefore, it is important to devise new methods to learn predictive models over datasets of permutations. However, maintaining probability distributions over the space of permutations is a hard task since there are

n

! permutations of

n

elements. The Fourier transform has been successfully generalized to functions over permutations. One of its main advantages in the context of probability distributions is that it compactly summarizes approximations to functions by discarding high order marginals information. In this paper, we present a method to learn a probability distribution that approximates the generating distribution of a given sample of permutations. In particular, this method learns the Fourier domain information representing this probability distribution.

This paper proposes an unsupervised approach that automatically detects and corrects a text containing multiple errors of both syntactic and semantic nature. The number of errors that can be corrected is equal to the number of correct words in the text. Error types include, but are not limited to: spelling errors, real-word spelling errors, typographical errors, unwanted words, missing words, prepositional errors, punctuation errors, and many of the grammatical errors (e.g., errors in agreement and verb formation).

In this study we design and implement an artificial emotional response algorithm using the Ortony, Clore and Collins theory in an effort to understand and better simulate the response of intelligent agents in the presence of emotional stress. We first develop a general model to outline a generic emotional agent behaviour. Agents are then socially connected and surrounded by objects, or other actors, that trigger various emotions. A case study is built using a basic hospital model where nurse servicing patients interact in various static and dynamic emotional scenarios. The simulated results show that increase in emotional stress leads to higher error rates in nurse task performance.

The main constrains for Wireless Sensor Network (WSN) are its limited energy and bandwidth. In industry, WSN deployed with massive node density produces lots of sensory traffic with redundancy. Accordingly, it decreases the network lifetime. In our proposed approach, we investigate the problem on energy-efficient routing for a WSN in a radio harsh environment. We propose a novel approach to create optimal routing paths by using Genetic Algorithm (GA) and Dijkstra’s algorithm performed at Base Station (BS). To demonstrate the feasibility of our approach, formal analysis and simulation results are presented.

Despite the growth in digitization of data, there are still many languages without sufficient corpora to achieve valid measures of semantic similarity. If it could be shown that manually-assigned similarity scores from one language can be transferred to another language, then semantic similarity values could be used for languages with fewer resources. We test an automatic word similarity measure based on second-order co-occurrences in the Google n-gram corpus, for English, German, and French. We show that the scores manually-assigned in the experiments of Rubenstein and Goodenough’s for 65 English word pairs can be transferred directly into German and French. We do this by conducting human evaluation experiments for French word pairs (and by using similarly produced scores for German). We show that the correlation between the automatically-assigned semantic similarity scores and the scores assigned by human evaluators is not very different when using the Rubenstein and Goodenough’s scores across language, compared to the language-specific scores.

The clustering of related words is crucial for a variety of Natural Language Processing applications. Many known techniques of word clustering use the context of a word to determine its meaning. Words which frequently appear in similar contexts are assumed to have similar meanings. Word clustering usually applies the weighting of contexts, based on some measure of their importance. One of the most popular measures is Pointwise Mutual Information. It increases the weight of contexts where a word appears regularly but other words do not, and decreases the weight of contexts where many words may appear. Essentially, it is unsupervised feature weighting. We present a method of supervised feature weighting. It identifies contexts shared by pairs of words known to be semantically related or unrelated, and then uses Pointwise Mutual Information to weight these contexts on how well they indicate closely related words. We use

Roget’s

Thesaurus

as a source of training and evaluation data. This work is as a step towards adding new terms to

Roget’s

Thesaurus

automatically, and doing so with high confidence.

This paper employs SPOT (Stream Projected Outlier deTector) as a prototype system for anomaly-based intrusion detection and evaluates its performance against other major methods. SPOT is capable of processing high-dimensional data streams and detecting novel attacks which exhibit abnormal behavior, making it a good candidate for network intrusion detection. This paper demonstrates SPOT is effective to distinguish between normal and abnormal processes in a UNIX System Call dataset.

Scenario based design (SBD) approach has been widely used to improve the user interface (UI) design of interactive systems. In this paper, the effectiveness of using the SBD approach is shown to improve the UI design of the Thunderbird email system. Firstly, an empirical evaluation of the system was performed based on the user comments. Then, a low fidelity prototype of the modified interfaces was developed. Furthermore, the new design interfaces were evaluated using two evaluation methods: a) GOMS keystroke level model was used to compare the efficiency of two interfaces b) Heuristic Evaluation of the system was performed using Nielsen’s usability heuristics. The evaluation results show that the efficiency of accomplishing important and most discussed tasks is improved significantly. Applying SBD approach on email systems is concluded as a promising trend to enhance usability.

Due to the rapidly increasing amount of biomedical literature, automatic processing of biomedical papers is extremely important. Named Entity Recognition (NER) in this type of writing has several difficulties. In this paper we present a system to find phenotype names in biomedical literature. The system is based on Metamap and makes use of the UMLS Metathesaurus and the Human Phenotype Ontology. From an initial basic system that uses only these preexisting tools, five rules that capture stylistic and linguistic properties of this type of literature are proposed to enhance the performance of our NER tool. The tool is tested on a small corpus and the results (precision 97.6% and recall 88.3%) demonstrate its performance.

Learning from data with severe class imbalance is difficult. Established solutions include: under-sampling, adjusting classification threshold, and using an ensemble. We examine the performance of combining these solutions to balance the sensitivity and specificity for binary classifications, and to reduce the MSE score for probability estimation.

We present a new tool: symbolic dynamical approach to the simulation of cognitive processes. Complex Auto-Adaptive System is a symbolic dynamical system implemented in a multi-agent system. We describe our methodology and prove our claims by presenting simulation experiments of the Stroop and Wason tasks. We then explain our research plan: the integration of an emotion model in our system, the implementation of a higher-level control organization, and the study of its application to cognitive agents in general.

Although propositional satisfiability (SAT) is NP-complete, state-of-the-art SAT solvers are able to solve large, practical instances. The concept of backdoors has been introduced to capture structural properties of instances. A backdoor is a set of variables that, if assigned correctly, leads to a polynomial-time solvable sub-problem. In this paper, we address the problem of finding all small backdoors, which is essential for studying value and variable ordering mistakes. We discuss our definition of sub-solvers and propose algorithms for finding backdoors. We experimentally compare our proposed algorithms to previous algorithms on structured and real-world instances. Our proposed algorithms improve over previous algorithms for finding backdoors in two ways. First, our algorithms often find smaller backdoors. Second, our algorithms often find a much larger number of backdoors.

Personalized Web search systems have been developed to tailor Web search to users’ needs based on their interests and preferences. A novel Normal Distribution Re-Weighting (NDRW) approach is proposed in this paper, which identifies and re-weights significant terms in vector-based personalization models in order to improve the personalization process. Machine learning approaches will be used to train the algorithm and discover optimal settings for the NDRW parameters. Correlating these parameters to features of the personalization model will allow this re-weighting process to become automatic.

Hierarchical problem solving, in terms of abstraction hierarchies or granular state spaces, is an effective way to structure state space for speeding up a search process. However, the problem of constructing and interpreting an abstraction hierarchy is still not fully addressed. In this paper, we propose a framework for constructing granular state spaces by applying results from granular computing and rough set theory. The framework is based on an addition of an information table to the original state space graph so that all the states grouped into the same abstract state are graphically and semantically close to each other.

Speech is a complex process that requires control and coordination of articulation, breathing, voicing, and prosody. Dysarthria is a manifestation of an inability to control and coordinate one or more of these aspects, which results in poorly articulated and hardly intelligible speech. Hence individuals with dysarthria are rarely understood by human listeners. In this paper, we compare and evaluate how well dysarthric speech can be recognized by an automatic speech recognition system (ASR) and naïve adult human listeners. The results show that despite the encouraging performance of ASR systems, and contrary to the claims in other studies, on average human listeners perform better in recognizing single-word dysarthric speech. In particular, the mean word recognition accuracy of speaker-adapted monophone ASR systems on stimuli produced by six dysarthric speakers is 68.39% while the mean percentage correct response of 14 naïve human listeners on the same speech is 79.78% as evaluated using single-word multiple-choice intelligibility test.

In this paper we have proposed a context-aware reputation-based trust model for multi-agent environments. Due to the lack of a general method for recognition and representation of context notion, we proposed a functional ontology of context for evaluating trust (FOCET) as the building block of our model. In addition, a computational reputation-based trust model based on this ontology is developed. Our model benefits from powerful reasoning facilities and the capability of adjusting the effect of context on trust assessment. Simulation results shows that an appropriate context weight results in the enhancement of the total profit in open systems.

Today, research on automatic text summarization challenges on readability factor as one of the most important aspects of summarizers’ performance. In this paper, we present Pazesh: a language-independent graph-based approach for increasing the readability of summaries while preserving the most important content. Pazesh accomplishes this task by constructing a special path of salient sentences which passes through topic centroid sentences. The results show that Pazesh compares approvingly with previously published results on benchmark datasets.

The evolution of artificial intelligence has followed our needs. At first, there was a need for the production of information, followed by the need to store digital data. Following the explosion in the amount of generated and stored data, we needed to find the information we require. The problem is now how to present this information to the user. In this paper we present ideas and research directions that we want to explore in order to develop a new approaches and methods for the synthetic presentation of objective information.

Automated thesaurus construction by collecting relations between lexical items (synonyms, antonyms, etc) has a long tradition in natural language processing. This has been done by exploiting dictionary structures or distributional context regularities (coocurrence, syntactic associations, or translation equivalents), in order to define measures of lexical similarity or relatedness. Dyvik had proposed to use aligned multilingual corpora and defines similar terms as terms that often share their translations. We evaluate the usefulness of this similarity for the extraction of synonyms, compared to the more widespread distributional approach.

Valuation algebras abstract a large number of formalisms for automated reasoning and enable the definition of generic inference procedures. Many of these formalisms provide some notions of solutions. Typical examples are satisfying assignments in constraint systems, models in logics or solutions to linear equation systems. Contrary to inference, there is no general algorithm to compute solutions in arbitrary valuation algebras. This paper states formal requirements for the presence of solutions and proposes a generic algorithm for solution construction based on the results of a previously executed inference scheme. We study the application of generic solution construction to semiring constraint systems, sparse linear systems and algebraic path problems and show that the proposed method generalizes various existing approaches for specific formalisms in the literature.

Word Sense Disambiguation has long been a central problem in computational linguistics. Word Sense Disambiguation is the ability to identify the meaning of words in context in a computational manner. Statistical and supervised approaches require a large amount of labeled resources as training datasets. In contradistinction to English, the Persian language has neither any semantically tagged corpus to aid machine learning approaches for Persian texts, nor any suitable parallel corpora. Yet due to the ever-increasing development of Persian pages in Wikipedia, this resource can act as a comparable corpus for English-Persian texts.

In this paper, we propose a cross-lingual approach to tagging the word senses in Persian texts. The new approach makes use of English sense disambiguators, the Wikipedia articles in both English and Persian, and a newly developed lexical ontology, FarsNet. It overcomes the lack of knowledge resources and NLP tools for the Persian language. We demonstrate the effectiveness of the proposed approach by comparing it to a direct sense disambiguation approach for Persian. The evaluation results indicate a comparable performance to the utilized English sense tagger.

Contrast sets have been shown to be a useful mechanism for describing differences between groups. A contrast set is a conjunction of attribute-value pairs that differ significantly in their distribution across groups. These groups are defined by a selected property that distinguishes one from the other (e.g customers who default on their mortgage versus those that don’t). In this paper, we propose a new search algorithm which uses a vertical approach for mining maximal contrast sets on categorical and quantitative data. We utilize a novel yet simple discretization technique, akin to simple binning, for continuous-valued attributes. Our experiments on real datasets demonstrate that our approach is more efficient than two previously proposed algorithms, and more effective in filtering interesting contrast sets.

Ontology classification is an essential reasoning task for ontology based systems. Tableau and resolution are two dominant types of reasoning procedures for ontology reasoning. Complex ontologies are often built on more expressive description logics and are usually highly cyclic. When reasoning complex ontologies, the both approaches may have difficulties in terms of reasoning results and performance, but for different ontology types. In this research, we investigate a hybrid reasoning approach, which will employ well-defined strategies to decompose and modify a complex ontology into subsets of ontologies based on capabilities of different reasoners, process the subsets with suitable individual reasoners, and combine such individual classification results into the overall classification result. The objective of our approach is to detect more subsumption relationships than individual reasoners for complex ontologies, and improve overall reasoning performance.

Subspace mapping methods aim at projecting high-dimensional data into a subspace where a specific objective function is optimized. Such dimension reduction allows the removal of collinear and irrelevant variables for creating informative visualizations and task-related data spaces. These specific and generally de-noised subspaces spaces enable machine learning methods to work more efficiently. We present a new and general subspace mapping method, Correlative Matrix Mapping (CMM), and evaluate its abilities for category-driven text organization by assessing neighborhood preservation, class coherence, and classification. This approach is evaluated for the challenging task of processing short and noisy documents.

This paper introduces AdaBoost Dynamic, an extension of AdaBoost.M1 algorithm by Freund and Shapire. In this extension we use different “weak” classifiers in subsequent iterations of the algorithm, instead of AdaBoost’s fixed base classifier. The algorithm is tested with various datasets from UCI database, and results show that the algorithm performs equally well as AdaBoost with the best possible base learner for a given dataset. This result therefore relieves a machine learning analyst from having to decide which base classifier to use.

The ability to efficiently perform probabilistic inference task is critical to large scale applications in statistics and artificial intelligence. Dramatic speedup might be achieved by appropriately mapping the current inference algorithms to the parallel framework. Parallel exact inference methods still suffer from exponential complexity in the worst case. Approximate inference methods have been parallelized and good speedup is achieved. In this paper, we focus on a variant of Belief Propagation algorithm. This variant has better convergent property and is provably convergent under certain conditions. We show that this method is amenable to coarse-grained parallelization and propose techniques to optimally parallelize it without sacrificing convergence. Experiments on a shared memory systems demonstrate that near-ideal speedup is achieved with reasonable scalability.

A widespread idea to attack ranking works by reducing it into a set of binary preferences and applying well studied classification techniques. The basic question addressed in this paper relates to whether an accurate classifier would transfer directly into a good ranker. In particular, we explore this reduction for subset ranking, which is based on optimization of DCG metric (Discounted Cumulated Gain), a standard position-sensitive performance measure. We propose a consistent reduction framework, guaranteeing that the minimal DCG regret is achievable by learning pairwise preferences assigned with importance weights. This fact allows us to further develop a novel upper bound on the DCG regret in terms of pairwise regrets. Empirical studies on benchmark datasets validate the proposed reduction approach with improved performance.

Software engineers need to be able to create, modify, and analyze knowledge stored in software artifacts. A significant amount of these artifacts contain natural language, like version control commit messages, source code comments, or bug reports. Integrated software development environments (IDEs) are widely used, but they are only concerned with structured software artifacts – they do not offer support for analyzing unstructured natural language and relating this knowledge with the source code. We present an integration of natural language processing capabilities into the Eclipse framework, a widely used software IDE. It allows to execute NLP analysis pipelines through the Semantic Assistants framework, a service-oriented architecture for brokering NLP services based on GATE. We demonstrate a number of semantic analysis services helpful in software engineering tasks, and evaluate one task in detail, the quality analysis of source code comments.

We consider how to plan optimally in a testbed, multiagent expedition (MAE), by centralized or distributed computation. As optimal planning in MAE is highly intractable, we investigate speedup through partial evaluation of a subset of plans whereby only the intended effect of a plan is evaluated when certain conditions hold. We apply this technique to centralized planning and demonstrate significant speedup in runtime while maintaining optimality. We investigate the technique in distributed planning and analyze the pitfalls.