Advances in Artificial Intelligence

This talk presents a synergistic methodology based on generalized stochastic Petri-nets (SPN) and neural nets for efficiently developing planning strategies. The SPN planning method generates global plans based on the states of the elements of the Universe of Discourse. Each plan includes all the possible conflict free planning paths for achieving the desirable goals under certain constraints occurred at the problem to be solved. The a neural network is used for searching the vectors of markings generated by the SPN reachability graph for the appropriate selection of plans. The SPN model presents high complexity issues, but at the same time offers to the synergic important features, such as stochastic modeling, synchronization, parallelism, concurrency and timing of events, valuable for developing plans under uncertainty. The neural network does contribute to the high complexity, but it offers learning capability to the synergy for future use. An example for coordinating two robotic arms under the constraints of time, space, and placement of the objects will be presented.

What patterns can we find in a bursty web traffic? On the web or on the internet graph itself? How about the distributions of galaxies in the sky, or the distribution of a company’s customers in geographical space? How long should we expect a nearest-neighbour search to take, when there are 100 attributes per patient or customer record? The traditional assumptions (uniformity, independence, Poisson arrivals, Gaussian distributions), often fail miserably. Should we give up trying to find patterns in such settings? Self-similarity, fractals and power laws are extremely successful in describing real datasets (coast-lines, rivers basins, stock-prices, brain-surfaces, communication-line noise, to name a few). We show some old and new successes, involving modeling of graph topologies (internet, web and social networks); modeling galaxy and video data; dimensionality reduction; and more.

In this work, we model speech samples with a two-sided generalized Gamma distribution and evaluate its efficiency for voice activity detection. Using a computationally inexpensive maximum likelihood approach, we employ the Bayesian Information Criterion for identifying the phoneme boundaries in noisy speech.

As the field of Intelligent Virtual Agents evolves and advances, an ever increasing number of functional and useful applications are presented. Intelligent Virtual Agents have become more realistic, intelligent and sociable, with apparent and substantial benefits to domains such as training, tutoring, simulation and entertainment. However, even though many end-users can enjoy these benefits today, the development of such applications is restricted to specialized research groups and companies. Obvious and difficult-to-overcome factors contribute to this. The inherent complexity of such applications results in increased theoretical and technical requirements to their development. Furthermore, Intelligent Virtual Agent systems today typically offer ad hoc, if any, design and development means that lack completeness and a general-purpose character. Significant efforts have been successfully made towards deriving globally accepted standards; nevertheless these mostly focus on communication between heterogeneous systems and not on design and development. In this paper, we present our current efforts towards a novel architecture for Intelligent Virtual Agents which is based on our previous work in the field and encompasses the full range of characteristics considered today as fundamental to achieving believable Intelligent Virtual Agent behaviour. In the spirit of enabling and easing application design and development, as well as facilitating further research, our architecture is tightly coupled with a behaviour specification language that uniformly covers all aspects and stages of the development process. We also present the key guidelines for a minimal but functional implementation, aimed in validation and experimentation.

In this paper a new maximum a posteriori (MAP) approach based on mixtures of multinomials is proposed for discovering probabilistic patterns in sequences. The main advantage of the method is the ability to bypass the problem of overlapping patterns in neighboring positions of sequences by using a Markov random field (MRF) prior. This model consists of two components, the first models the pattern and the second the background. The Expectation-Maximization (EM) algorithm is used to estimate the model parameters and provides closed form updates. Special care is also taken to overcome the known dependence of the EM algorithm to initialization. This is done by applying an adaptive clustering scheme based on the

k

-means algorithm in order to produce good initial values for the pattern multinomial model. Experiments with artificial sets of sequences show that the proposed approach discovers qualitatively better patterns, in comparison with maximum likelihood (ML) and Gibbs sampling (GS) approaches.

A hardware architecture is presented, which accelerates the per- formance of intelligent applications that are based on logic programming. The logic programs are mapped on hardware and more precisely on FPGAs (Field Programmable Gate Array). Since logic programs may easily be transformed into an equivalent Attribute Grammar (AG), the underlying model of implementing an embedded system for the aforementioned applications can be that of an AG evaluator. Previous attempts to the same problem were based on the use of two separate components. An FPGA was used for mapping the inference engine and a conventional RISC microprocessor for mapping the unification mechanism and user defined additional semantics. In this paper a new architecture is presented, in order to drastically reduce the number of the required processing elements by a factor of

n

(length of input string). This fact and the fact of using, for the inference engine, an extension of the most efficient parsing algorithm, allowed us to use only one component i.e. a single FPGA board, eliminating the need for an additional external RISC microprocessor, since we have embedded two “PicoBlaze” Soft Processors into the FPGA. The proposed architecture is suitable for embedded system applications where low cost, portability and low power consumption is of crucial importance. Our approach was tested with numerous examples in order to establish the performance improvement over previous attempts.

Decision making is widely considered as a fundamental organizational activity that comprises a series of knowledge representation and processing tasks. Admitting that the quality of a decision depends on the quality of the knowledge used to make it, it is argued that the enhancement of the decision making efficiency and effectiveness is strongly related to the appropriate exploitation of all possible organizational knowledge resources. Taking the above remarks into account, this paper presents a multidisciplinary approach for capturing the organizational knowledge in order to augment teamwork in terms of knowledge elicitation, sharing and construction, thus enhancing decision making quality. Based on a properly defined ontology model, our approach is supported by a web-based tool that serves as a forum of reciprocal knowledge exchange, conveyed through structured argumentative discourses, the ultimate aim being to support the related decision making process. The related knowledge is represented through a Discourse Graph, which is structured and evaluated according to the knowledge domain of the problem under consideration.

The market clearing prices in deregulated electricity markets are volatile. Good market clearing price forecasting will help producers and consumers to prepare their corresponding bidding strategies so as to maximize their profits. Market clearing price prediction is a difficult task since bidding strategies used by market participants are complicated and various uncertainties interact in an intricate way. This paper proposes an adaptively trained neural network to forecast the 24 day-ahead market-clearing prices. The adaptive training mechanism includes a feedback process that allows the artificial neural network to learn from its mistakes and correct its output by adjusting its architecture as new data becomes available. The methodology is applied to the California power market and the results prove the efficiency and practicality of the proposed method.

A random signal-based learning merged with simulated annealing (SARSL), which is serial algorithm, has been considered by the authors. But the serial nature of SARSL degrades its performance as the complexity of the search space is increasing. To solve this problem, this paper proposes a population structure of SARSL (PSARSL) which enables multi-point search. Moreover, adaptive partitioning method (APM) is used to reduce the optimization time. The validity of the proposed algorithm is conformed by applying it to a simple test function example and a general version of fuzzy controller design.

This paper is concerned with cascade fuzzy neural networks and its optimization. These networks come with sound and transparent logic characteristics by being developed with the aid of AND and OR fuzzy neurons and subsequently logic processors (LPs). We discuss main functional properties of the model and relate them to its form of cascade type of systems formed as a stack of LPs. The structure of the network that deals with a selection of a subset of input variables and their distribution across the individual LPs is optimized with the use of genetic algorithms (GA). We discuss random signal-based learning (RSL), a local search technique, aimed at further refinement of the connections of the neurons (GA-RSL). We elaborate on the interpretation aspects of the network and show how this leads to a Boolean or multi-valued logic description of the experimental data. Two kinds of standard data sets are discussed with respect to the performance of the constructed networks and their interpretability.

Video broadcast series like news or magazine broadcasts usually expose a strong temporal structure, along with a characteristic audio-visual appearance. This results in frequent patterns occurring in the video signal. We propose an algorithm for the automatic detection of such patterns that exploits the video’s self-similarity induced by the patterns. The approach is applied to the problem of anchor shot detection, but can also be used for other related purposes. Tests on real-world video data show that it is possible with our method to detect anchor shots fully automatically with high reliability.

This paper presents a content-based approach to spam detection based on low-level information. Instead of the traditional ’bag of words’ representation, we use a ’bag of character

n

-grams’ representation which avoids the sparse data problem that arises in

n

-grams on the word-level. Moreover, it is language-independent and does not require any lemmatizer or ’deep’ text preprocessing. Based on experiments on Ling-Spam corpus we evaluate the proposed representation in combination with support vector machines. Both binary and term-frequency representations achieve high precision rates while maintaining recall on equally high level, which is a crucial factor for anti-spam filters, a cost sensitive application.

The intermittent nature of the wind creates significant uncertainty in the operation of power systems with increased wind power penetration. Con- siderable efforts have been made for the accurate prediction of the wind power using either statistical or physical models. In this paper, a method based on Artificial Neural Network (ANN) is proposed in order to improve the predictions of an existing neuro-fuzzy wind power forecasting model taking into account the evaluation results from the use of this wind power forecasting tool. Thus, an improved wind power forecasting is achieved and a better estimation of the confidence interval of the proposed model is provided.

This paper presents a new bidding strategy for continuous double auctions (CDA) designed for Mertacor, a successful trading agent, which won the first price in the “travel game” of Trading Agent Competition (TAC) for 2005. TAC provides a realistic benchmarking environment in which various travel commodities are offered in simultaneous online auctions. Among these, entertainment tickets are traded in CDA. The latter, represent the most dynamic part of the TAC game, in which agents are both sellers and buyers. In a CDA many uncertainty factors are introduced, because prices are constantly changing during the game and price fluctuations are hard to be predicted. In order to deal with these factors of uncertainty we have designed a strategy based on achieving a pre-defined long-term profit. This preserves the bidding attitude of our agent and shows flexibility in changes of the environment. We finally present and discuss the results of TAC-05, as well as an analysis of agents performance in the entertainment auctions.

The leap from decision support to autonomous systems has often raised a number of issues, namely system safety, soundness and security. Depending on the field of application, these issues can either be easily overcome or even hinder progress. In the case of Supply Chain Management (SCM), where system performance implies loss or profit, these issues are of high importance. SCM environments are often dynamic markets providing incomplete information, therefore demanding intelligent solutions which can adhere to environment rules, perceive variations, and act in order to achieve maximum revenue. Advancing on the way such autonomous solutions deal with the SCM process, we have built a robust, highly-adaptable and easily-configurable mechanism for efficiently dealing with all SCM facets, from material procurement and inventory management to goods production and shipment. Our agent has been crash-tested in one of the most challenging SCM environments, the trading agent competition SCM game and has proven capable of providing advanced SCM solutions on behalf of its owner. This paper introduces

Mertacor

and its main architectural primitives, provides an overview of the TAC SCM environment, and discusses

Mertacor

’s performance.

Probabilistic Latent Semantic Indexing (PLSI) is a statistical technique for automatic document indexing. A novel method is proposed for updating PLSI when new documents arrive. The proposed method adds incrementally the words of any new document in the term-document matrix and derives the updating equations for the probability of terms given the class (i.e. latent) variables and the probability of documents given the latent variables. The performance of the proposed method is compared to that of the folding-in algorithm, which is an inexpensive, but potentially inaccurate updating method. It is demonstrated that the proposed updating algorithm outperforms the folding-in method with respect to the mean squared error between the aforementioned probabilities as they are estimated by the two updating methods and the original non-adaptive PLSI algorithm.

Parametric models such as linear regression can provide useful, interpretable descriptions of simple structure in data. However, sometimes such simple structure does not extend across an entire data set and may instead be confined more locally within subsets of the data. Nonparametric regression typically involves local averaging. In this study, local averaging estimator is coupled with a machine learning technique – boosting. In more detail, we propose a technique of local boosting of decision stumps. We performed a comparison with other well known methods and ensembles, on standard benchmark datasets and the performance of the proposed technique was greater in most cases.

We present,

Mine Time

, a tool that supports discovery over time series data.

Mine Time

is realized by the introduction of novel algorithmic processes, which support assessment of coherence and similarity across timeseries data. The innovation comes from the inclusion of specific ‘control’ operations in the elaborated time-series matching metric. The final outcome is the clustering of time-series into similar-groups. Clustering is performed via the appropriate customization of a phylogeny-based clustering algorithm and tool. We demonstrate

Mine Time

via two experiments.

Several agent frameworks have been proposed for developing intelligent software agents and multi-agent systems that are able to perform in dynamic environments. These frameworks and architectures exploit specific reasoning tasks (such as option selection, desire filtering, plan elaboration and means-end reasoning) that support agents to react, deliberate and/or interact/cooperate with other agents. Such reasoning tasks are realized by means of specific modules that agents may trigger according to circumstances, switching their behaviour between predefined discrete behavioural modes. This paper presents the facilities provided by the non-layered BDI-architecture of IC

agent

for supporting performance in dynamic and unpredictable multi-agent environments through efficient balancing between behavioural modes in a continuous space. This space is circumscribed by the purely (individual) reactive, the purely (individual) deliberative and the social deliberative behavioural modes. In a greater extend than existing frameworks; IC

agent

relates agent’s flexible behaviour to cognition and sociability, supporting the management of plans constructed by the agent’s mental and domain actions in a coordinated manner.

This paper describes the application of Decision Tress (DTs) in order to specify the most critical location and the rate of series compensation in order to increase power system loading margin. The proposed methodology is applied to a projected model of the Hellenic interconnected system in several system configurations. Investigation of the best system operating point to create the DTs, the effect of attributes number and type on the DTs size and quality are discussed in order to reach the final DTs parameters that lead to the construction of the best DTs for the determination of optimal series compensation location and rate. Finally, the results obtained for several (N-1) contingencies examined are presented.

In this paper, a new genetic clustering algorithm called IHGA-clustering is proposed to deal with the clustering problem under the criterion of minimum sum of squares clustering. In IHGA-clustering, DHB operation is developed to improve the individual and accelerate the convergence speed, and partition-mergence mutation operation is designed to reassign objects among different clusters. Equipped with these two components, IHGA-clustering can stably output the proper result. Its superiority over HGA-clustering, GKA, and KGA-clustering is extensively demonstrated for experimental data sets.

We present a named-entity recognizer for Greek person names and temporal expressions. For temporal expressions, it relies on semi- automatically produced patterns. For person names, it employs two Support Vector Machines, that scan the input text in two passes, and active learning, which reduces the human annotation effort during training.

During the last years, computer vision-based diagnostic systems have been used in several hospitals and dermatology clinics, aiming mostly at the early detection of malignant melanoma tumor, which is among the most frequent types of skin cancer, versus other types of non-malignant cutaneous diseases. In this paper we discuss intelligent techniques for the segmentation and classification of pigmented skin lesions in such dermatological images. A local thresholding algorithm is proposed for skin lesion separation and border, texture and color based features, are then extracted from the digital images. Extracted features are used to construct a classification module based on Support Vector Machines (SVM) for the recognition of malignant melanoma versus dysplastic nevus.

The current work addresses the development of cognitive abilities in artificial organisms. In the proposed approach, neural network-based agent structures are employed to represent distinct brain areas. We introduce a Hierarchical Cooperative CoEvolutionary (HCCE) approach to design autonomous, yet collaborating agents. Thus, partial brain models consisting of many substructures can be designed. Replication of lesion studies is used as a means to increase reliability of brain model, highlighting the distinct roles of agents. The proposed approach effectively designs cooperating agents by considering the desired pre- and post- lesion performance of the model. In order to verify and assess the implemented model, the latter is embedded in a robotic platform to facilitate its behavioral capabilities.

The present paper discusses the issue of enhancing classification performance by means other than improving the ability of certain Machine Learning algorithms to construct a precise classification model. On the contrary, we approach this significant problem from the scope of an extended coding of training data. More specifically, our method attempts to generate more features in order to reveal the hidden aspects of the domain, modeled by the available training examples. We propose a novel feature construction algorithm, based on the ability of Bayesian networks to represent the conditional independence assumptions of a set of features, thus projecting relational attributes which are not always obvious to a classifier when presented in their original format. The augmented set of features results in a significant increase in terms of classification performance, a fact that is depicted to a plethora of machine learning domains (i.e. data sets from the UCI ML repository and the Artificial Intelligence group) using a variety of classifiers, based on different theoretical backgrounds.

Traditionally, musical instrument recognition is mainly based on frequency domain analysis (sinusoidal analysis, cepstral coefficients) and shape analysis to extract a set of various features. Instruments are usually classified using k-NN classifiers, HMM, Kohonen SOM and Neural Networks. In this work, we describe a system for the recognition of musical instruments from isolated notes. We are introducing the use of a Time Encoded Signal Processing method to produce simple matrices from complex sound waveforms, for instrument note encoding and recognition. These matrices are presented to a Fast Artificial Neural Network (FANN) to perform instrument recognition with promising results in organ classification and reduced computational cost. The evaluation material consists of 470 tones from 19 musical instruments synthesized with 5 wide used synthesizers (Microsoft Synth, Creative SB Live! Synth, Yamaha VL-70m Tone Generator, Edirol Soft-Synth, Kontakt Player) and 84 isolated notes from 20 western orchestral instruments (Iowa University Database).

This paper reports on the implementation of a rule system, called O-DEVICE, for reasoning about OWL instances using deductive rules. O-DEVICE exploits the rule language of the CLIPS production rule system and transforms OWL ontologies into an object-oriented schema of COOL. During the transformation procedure, OWL classes are mapped to COOL classes, OWL properties to class slots and OWL instances to COOL objects. The purpose of this transformation is twofold: a) to exploit the advantages of the object-oriented representation and access all the properties of instances in one step, since properties are encapsulated inside resource objects; b) to be able to use a deductive object-oriented rule language for querying and creating maintainable views of OWL instances, which operates over the object-oriented schema of CLIPS, and c) to answer queries faster, since the implied relationships due to the rich OWL semantics have been pre-computed. The deductive rules are compiled into CLIPS production rules. The rich open-world semantics of OWL are partly handled by the incremental transformation procedure and partly by the rule compilation procedure.

The extraction of information from a source containing term-classified objects is plagued with uncertainty, due, among other things, to the possible incompleteness of the source index. To overcome this incompleteness, the study proposes to expand the index of the source, in a way that is as reasonable as possible with respect to the original classification of objects. By equating reasonableness with logical implication, the sought expansion turns out to be an explanation of the index, captured by abduction. We study the general problem of query evaluation on the extended information source, providing a polynomial time algorithm which tackles the general case, in which no hypothesis is made on the structure of the taxonomy. We then specialize the algorithm for two well-know structures: DAGs and trees, showing that each specialization results in a more efficient query evaluation.

The article presents in detail a model for the assessment of feature weight values in context of inductive machine learning. Weight assessment is done based on learned knowledge and can not be used to assess feature values prior to learning. The model is based on Ackoff’s theory of behavioral communication. The model is also used to assess rule value importance. We present model heuristics and present a simple application based on the “play” vs. “not play” golf application. Implications about decision making modeling are discussed.

In the present work we study the applicability of Support Vector Machines (SVMs) on the phoneme recognition task. Specifically, the Least Squares version of the algorithm (LS-SVM) is employed in recognition of the Greek phonemes in the framework of telephone-driven voice-enabled information service. The N-best candidate phonemes are identified and consequently feed to the speech and language recognition components. In a comparative evaluation of various classification methods, the SVM-based phoneme recognizer demonstrated a superior performance. Recognition rate of 74.2% was achieved from the N-best list, for N=5, prior to applying the language model.

Multiple Classifier systems have been developed in order to improve classification accuracy using methodologies for effective classifier combination. Classical approaches use heuristics, statistical tests, or a meta-learning level in order to find out the optimal combination function. We study this problem from a Reinforcement Learning perspective. In our modeling, an agent tries to learn the best policy for selecting classifiers by exploring a state space and considering a future cumulative reward from the environment. We evaluate our approach by comparing with state-of-the-art combination methods and obtain very promising results.

The role and importance of methods for lexical knowledge elicitation in the area of multilingual information processing, including machine translation, computer-aided translation and cross-lingual information retrieval is undisputable. The usefulness of such methods becomes even more apparent in cases of language pairs where no appropriate digital language resources exist. This paper presents encouraging experimental results in automatically eliciting bilingual lexica out of Greek-Turkish parallel corpora, consisting of international organizations’ documents available in English, Greek and Turkish, in an attempt to aid multilingual document processing involving these languages.

In this paper feed-forward neural networks are introduced where hidden units employ orthogonal Hermite polynomials for their activation functions. The proposed neural networks have some interesting properties: (i) the basis functions are invariant under the Fourier transform, subject only to a change of scale, and (ii) the basis functions are the eigenstates of the quantum harmonic oscillator, and stem from the solution of Schrödinger’s diffusion equation. The proposed neural networks demonstrate the particle-wave nature of information and can be used in nonparametric estimation. Possible applications of neural networks with Hermite basis functions include system modelling and image processing.

Coalition formation is an important area of research in multi-agent systems. Computing optimal coalition structures for a large number of agents is an important problem in coalition formation but has received little attention in the literature. Previous studies assume that each coalition value is known a priori. This assumption is impractical in real world settings. Furthermore, the problem of finding coalition values become intractable for even a relatively small number of agents. This work proposes a distributed branch-and-bound algorithm for computing optimal coalition structures in linear production domain, where each coalition value is not known a priori. The common goal of the agents is to maximize the system’s profit. In our algorithm, agents perform two tasks:

i

) deliberate profitable coalitions, and

ii

) cooperatively compute optimal coalition structures. We show that our algorithm outperforms exhaustive search in generating optimal coalition structure in terms of elapses time and number of coalition structures generated.

The innovative feature of the system presented in this paper is the use of pattern-matching techniques to retrieve translations resulting in a flexible, language-independent approach, which employs a limited amount of explicit a priori linguistic knowledge. Furthermore, while all state-of-the-art corpus-based approaches to Machine Translation (MT) rely on bitexts, this system relies on extensive target language monolingual corpora. The translation process distinguishes three phases: 1) pre-processing with ‘light’ rule and statisticsbased NLP techniques 2) search & retrieval, 3) synthesising. At Phase 1, the source language sentence is mapped onto a lemma-to-lemma translated string. This string then forms the input to the search algorithm, which retrieves similar sentences from the corpus (Phase 2). This retrieval process is performed iteratively at increasing levels of detail, until the best match is detected. The best retrieved sentence is sent to the synthesising algorithm (Phase 3), which handles phenomena such as agreement.

Bayesian Networks are proven to be a comprehensive model to describe causal relationships among domain attributes with probabilistic measure of appropriate conditional dependency. However, depending on task and context, many attributes of the model might not be relevant. If a network has been learned across multiple contexts then all uncovered conditional dependencies are averaged over all contexts and cannot guarantee high predictive accuracy when applied to a concrete case. We are considering a context as a set of contextual attributes, which are not directly effect probability distribution of the target attributes, but they effect on a “relevance” of the predictive attributes towards target attributes. In this paper we use the Bayesian Metanetwork vision to model such context-sensitive feature relevance. Such model assumes that the relevance of predictive attributes in a Bayesian network might be a random attribute itself and it provides a tool to reason based not only on probabilities of predictive attributes but also on their relevancies. According to this model, the evidence observed about contextual attributes is used to extract a relevant substructure from a Bayesian network model and then the predictive attributes evidence is used to reason about probability distribution of the target attribute in the extracted sub-network. We provide the basic architecture for such Bayesian Metanetwork, basic reasoning formalism and some examples.

The prediction of the translation initiation site (TIS) in a genomic sequence is an important issue in biological research. Several methods have been proposed to deal with it. However, it is still an open problem. In this paper we follow an approach consisting of a number of steps in order to increase TIS prediction accuracy. First, all the sequences are scanned and the candidate TISs are detected. These sites are grouped according to the length of the sequence upstream and downstream them and a number of features is generated for each one. The features are evaluated among the instances of every group and a number of the top ranked ones are selected for building a classifier. A new instance is assigned to a group and is classified by the corresponding classifier. We experiment with various feature sets and classification algorithms, compare with alternative methods and draw important conclusions.

As is widely known, the popular Black & Scholes model for option pricing suffers from systematic biases, as it relies on several highly questionable assumptions. In this paper we study the ability of neural networks (MLPs) in pricing call options on the S&P 500 index; in particular we investigate the effect of the hidden neurons in the in- and out-of-sample pricing. We modify the Black & Scholes model given the price of an option based on the no-arbitrage value of a forward contract, written on the same underlying asset, and we derive a modified formula that can be used for our purpose. Instead of using the standard backpropagation training algorithm we replace it with the Levenberg-Marquardt approach. By modifying the objective function of the neural network, we focus the learning process on more interesting areas of the implied volatility surface. The results from this transformation are encouraging.

The Relevance Vector Machine(RVM) is a widely accepted Bayesian model commonly used for regression and classification tasks. In this paper we propose a multikernel version of the RVM and present an alternative inference algorithm based on Fourier domain computation to solve this model for large scale problems, e.g. images. We then apply the proposed method to the object detection problem with promising results.

In this paper we present an artificial cortical network, inspired by the Human Visual System (HVS), which extracts the salient contours in color images. Similarly to the primary visual cortex, the network consists of orientation hypercolumns. Lateral connections between the hypercolumns are modeled by a new connection pattern based on co-exponentiality. The initial color edges of the image are extracted in a way inspired by the double-opponent cells of the HVS. These edges are inputs to the network, which outputs the salient contours based on the local interactions between the hypercolumns. The proposed network was tested on real color images and displayed promising performance, with execution times small enough even for a conventional personal computer.

Modern Power Systems often operate close to their stability limits in order to meet the continuously growing demand, due to the difficulties in expanding the generation and transmission system. An effective way to face power system contingencies that can lead to instability is load shedding. In this paper we propose a method to assess the dynamic performance of the Greek mainland Power System and to propose a load shedding scheme in order to maintain voltage stability under various loading conditions and operating states in the presence of critical contingencies including outages of one or more generating units in the south part of the system. A Self Organizing Map is utilized in order to classify the Load profiles of the Power System. With a decision tree the dynamic performance of each class is assessed. The classification of Load Profiles by the SOM, provide the load shedding scheme.

This paper reports on the results of an ongoing project for the development of a platform for e-Learning, which automatically constructs curricula based on available educational resources and the learners needs and abilities. The system under development, called PASER (Planner for the Automatic Synthesis of Educational Resources), uses an automated planner, which given the initial state of the problem (learner’s profile, preferences, needs and abilities), the available actions (study an educational resource, take an exam, join an e-learning course, etc.) and the goals (obtain a certificate, learn a subject, acquire a skill, etc.) constructs a complete educational curriculum that achieves the goals. PASER is compliant with the evolving educational metadata standards that describe learning resources (LOM), content packaging (CP), educational objectives (RDCEO) and learner related information (LIP).

This paper introduces quantitative measurements/metrics of qualitative entertainment features within computer game environments and proposes artificial intelligence (AI) techniques for optimizing entertainment in such interactive systems. A human-verified metric of interest (i.e. player entertainment in real-time) for predator/prey games and a neuro-evolution on-line learning (i.e. during play) approach have already been reported in the literature to serve this purpose. In this paper, an alternative quantitative approach to entertainment modeling based on psychological studies in the field of computer games is introduced and a comparative study of the two approaches is presented. Artificial neural networks (ANNs) and fuzzy ANNs are used to model player satisfaction (interest) in real-time and investigate quantitatively how the qualitative factors of

challenge

and

curiosity

contribute to human entertainment. We demonstrate that appropriate non-extreme levels of challenge and curiosity generate high values of entertainment and we discuss the extensibility of the approach to other genres of digital entertainment and edutainment.

In this paper we are introducing the employment of features extracted from Fujisaki’s parameterization of pitch contour for the task of emotion recognition from speech. In evaluating the proposed features we have trained a decision tree inducer as well as the instance based learning algorithm. The datasets utilized for training the classification models, were extracted from two emotional speech databases. Fujisaki’s parameters benefited all prediction models with an average raise of 9,52% in the total accuracy.

In this paper, a two-class pattern recognition problem is studied, namely the automatic detection of speech disorders such as vocal fold paralysis and edema by processing the speech signal recorded from patients affected by the aforementioned pathologies as well as speakers unaffected by these pathologies. The data used were extracted from the Massachusetts Eye and Ear Infirmary database of disordered speech. The linear prediction coefficients are used as input to the pattern recognition problem. Two techniques are developed. The first technique is an optimal linear classifier design, while the second one is based on the dual-space linear discriminant analysis. Two experiments were conducted in order to assess the performance of the techniques developed namely the detection of vocal fold paralysis for male speakers and the detection of vocal fold edema for female speakers. Receiver operating characteristic curves are presented. Long-term mean feature vectors are proven very efficient in detecting the voice disorders yielding a probability of detection that may approach 100% for a probability of false alarm equal to 9.52%.

The selection of the winding material in power transformers is an important task, since it has significant impact on the transformer manufacturing cost. This winding material selection has to be checked in every transformer design, which means that for each design, there is a need to optimize the transformer twice and afterwards to select the most economical design. In this paper, an Artificial Neural Network (ANN) is proposed for the selection of the winding material in power transformers, which significantly contributes in the reduction of the effort needed in the transformer design. The proposed ANN architecture provides 94.7% classification success rate on the test set. Consequently, this method is very suitable for industrial use because of its accuracy and implementation speed.

A multi-layered biomedical literature mining approach is presented aiming to the discovery of gene-gene correlations and the construction of respective

gene networks

. Utilization of the

Trie

-memory data structure enables efficient manipulation of different gene nomenclatures. The whole approach is coupled with a texts (biomedical abstracts)

classification

method. Experimental validation and evaluation results show the rationality, efficiency and reliability of the approach.

Autonomy is a crucial property of an artificial agent. The type of representational structures and the role they play in the preservation of an agent’s autonomy are pointed out. A framework of self-organised Peircean semiotic processes is introduced and it is then used to demonstrate the emergence of grounded representational structures in agents interacting with their environment.

There exist many approaches to agent-based conflict resolution which adopts argumentation as their underlying conflict resolution machinery. In most argumentation systems, the credibility of argument sources plays a minimal role. This paper focuses on combining credibility of sources in a source sensitive argumentation.

This paper presents the main characteristics of Logic Models Creator (LMC). LMC is a new educational environment for young students to build and explore logic models in graphical form. LMC allows visual representation of logic models using IF/THEN/ELSE constructs. In this paper we provide an overview of LMC architecture and discuss briefly an example of use of LMC. As discussed, LMC users in the reported case study managed to achieve effective communication and task evaluation during exploration of problems involving decision making.

One of the crucial tasks towards the realization of the Semantic Web vision is the efficient encoding of human knowledge in ontologies. The proper maintenance of these, usually large, structures and, in particular, their adaptation to new knowledge (ontology evolution) is one of the most challenging problems in current Semantic Web research. In this paper, we uncover a certain gap in current ontology evolution approaches and propose a novel research path based on belief change. We present some ideas in this direction and argue that our approach introduces an interesting new dimension to the problem that is likely to find important applications in the future.

In this paper, we propose a novel holistic methodology for keyword search in historical typewritten documents combining synthetic data and user’s feedback. The holistic approach treats the word as a single entity and entails the recognition of the whole word rather than of individual characters. Our aim is to search for keywords typed by the user in a large collection of digitized typewritten historical documents. The proposed method is based on: (i) creation of synthetic image words; (ii) word segmentation using dynamic parameters; (iii) efficient hybrid feature extraction for each image word and (iv) a retrieval procedure that is optimized by user’s feedback. Experimental results prove the efficiency of the proposed approach.

Image fingerprinting systems aim to extract unique and robust image descriptors (in analogy to human fingerprints). They search for images that are not only perceptually similar but replicas of an image generated through mild image processing operations. In this paper, we examine the use of color descriptors based on a 24-color quantized palette for image fingerprinting. Comparisons are provided between different similarity measures methods as well as regarding the use of color-only and spatial chromatic histograms.

In this paper a method for feature selection and classification of email spam messages is presented. The selection of features is performed in two steps: The selection is performed by measuring their entropy and a fine-tuning selection is implemented using a genetic algorithm. In the classification process, a Radial Basis Function Network is used to ensure robust classification rate even in case of complex cluster structure. The proposed method shows that, when using a two-level feature selection, a better accuracy is achieved than using one-stage selection. Also, the use of a lemmatizer or a stop-word list gives minimal classification improvement. The proposed method achieves 96-97% average accuracy when using only 20 features out of 15000.

This work studies the problem of violence detection in audio data, which can be used for automated content rating. We employ some popular frame-level audio features both from the time and frequency domain. Afterwards, several statistics of the calculated feature sequences are fed as input to a Support Vector Machine classifier, which decides about the segment content with respect to violence. The presented experimental results verify the validity of the approach and exhibit a better performance than the other known approaches.

This paper addresses the problem of text classification in high dimensionality spaces by applying linear weight updating classifiers that have been highly studied in the domain of machine learning. Our experimental results are based on the Winnow family of algorithms that are simple to implement and efficient in terms of computation time and storage requirements. We applied an exponential multiplication function to weight updates and we experimentally calculated the optimal values of the learning rate and the separating surface parameters. Our results are at the level of the best results that were reported on the family of linear algorithms and perform nearly as well as the top performing methodologies in the literature.

We use decision trees and genetic algorithms to analyze the academic performance of students throughout an academic year at a distance learning university. Based on the accuracy of the generated rules, and on cross-examinations of various groups of the same student population, we surprisingly observe that students’ performance is clustered around tutors.

Traditional document clustering techniques rely on single-term analysis, such as the widely used Vector Space Model. However, recent approaches have emerged that are based on Graph Models and provide a more detailed description of document properties. In this work we present a novel Significance-based Graph Model for Web documents that introduces a sophisticated graph weighting method, based on significance evaluation of graph elements. We also define an associated similarity measure based on the maximum common subgraph between the graphs of the corresponding web documents. Experimental results on artificial and real document collections using well-known clustering algorithms indicate the effectiveness of the proposed approach.

We present a combined clinico-genomic knowledge discovery (CGKD) process suited for linking gene-expression (microarray) and clinical patient data. The process present a multi-strategy mining approach realized by the smooth integration of three distinct data-mining components: clustering (based on a discretized k-means approach), association rules mining, and feature-selection for selecting discrimant genes. The proposed CGKD process is applied on a real-world gene-expression profiling study (i.e., clinical outcome of breast cancer patients). Assessment of the results demonstrates the rationality and reliability of the approach.

This paper describes an ontology data service (ODS) for supporting Search and Rescue (SaR) operations. The ontological model represents various aspects of the command, communication, and organisational structure of the SaR forces and the deployment and progress of a SaR operation. Furthermore, the ontology supports the semantic indexing of multimedia documents in the context of SaR processes and activities. This ODS supports a semantically-enhanced information and communication system for SaR forces. Modelling the spatio-temporal aspects of an operation in alignment with possibly-unreliable information automatically extracted from multimedia objects, introduces a number of challenges for the field of knowledge representation and reasoning.

Defeasible reasoning is a rule-based approach for efficient reasoning with incomplete and conflicting information. Nevertheless, it is based on solid mathematical formulations and is not fully comprehensible by end users, who often need graphical trace and explanation mechanisms for the derived conclusions. Directed graphs (or digraphs) can assist in this affair, but their applicability is balanced by the fact that it is difficult to associate data of a variety of types with the nodes and the connections in the graph. In this paper we try to utilize digraphs in the graphical representation of defeasible rules, by exploiting their expressiveness, but also trying to counter their major disadvantage, by defining multiple node and connection types.

We present a novel technique for efficient Content Based Peer to Peer Image Retrieval (CBP2PIR) that employs a Content Addressable Network (CAN). A two-stage color histogram based method is described. The first stage defines mapping into the CAN by use of a single fuzzy histogram; while the second stage completes the image retrieval process through a spatially-biased histogram. The proposed system is completely decentralized, non-flooding, and promises high image recall, while minimizing network traffic.

This paper explores the effectiveness of machine learning techniques in detecting firms that issue fraudulent financial statements (FFS) and deals with the identification of factors associated to FFS. To this end, a number of experiments have been conducted using representative learning algorithms, which were trained using a data set of 164 fraud and non-fraud Greek firms in the recent period 2001-2002. This study indicates that a decision tree can be successfully used in the identification of FFS and underline the importance of financial ratios.

The objective of this study is to investigate the discrimination of benign from malignant breast lesions using: the linear, the feedforward neural network, the k-nearest neighbor and the boosting classifiers. Nuclear morphometric parameters from cytological smears taken by Fine Needle Aspiration (FNA) of the breast, have been measured from

193

patients. These parameters undergo an appropriate transformation and then, the classifiers are performed on the raw and on the transformed data. The results show that in terms of the raw data set all classifiers exhibit almost the same performance (overall accuracy ≡ 87%), Thus the linear classifier suffices for the discrimination of the present problem. Also, based on the previous results, one can conjecture that the use of these classifiers combined with image morphometry and statistical techniques for feature transformation, may offer useful information towards the improvement of the diagnostic accuracy of breast FNA.

This paper proposes and compares several data fusion techniques for robot navigation. The fusion techniques investigated here are several topologies of the Kalman filter. The problem that had been simulated is the navigation of a robot carrying two sensors, one Global Positioning System (GPS) and one Inertial Navigation System (INS). For each of the above topologies, the statistic error and its, mean value, variance and standard deviation were examined.

This paper describes two simple techniques that can greatly improve navigation and motion control of nonholonomic robots based on range sensor data. The first technique enhances sensory information by re-using recent sensor data through coordinate transformation, whereas the second compensates for errors due to long control cycle times by forward projection through the kinematic model of the robot. Both techniques have been succesfully tested on a Nomad 200 mobile robot.

Let

G

(

V

,

E

) denote an undirected graph,

V

and

E

being the sets of its nodes and edges, respectively. A

matching

in

G

(

V

,

E

) is a subset of edges with no common endpoints. Finding a matching of maximum cardinality constitutes the maximum cardinality matching (MCM) problem. For a thorough theoretical discussion we refer to [6]. The MCM problem is of specific interest from a Constraint Programming (CP) point of view because it can model several logical constraints (predicates) like the

all_different

and the

symmetric all_different

predicates [7]. Thus, the definition of a maximum cardinality matching constraint provides a framework encompassing other predicates. Along this line of research, we define a global constraint with respect to the MCM and address the issue of consistency. Establishing hyper-arc consistency implies the identification of edges that cannot participate in any maximum cardinality matching. Evidently, this issue (also called

filtering

) is related to the methods developed for solving the problem. Solving this problem for bipartite graphs was common knowledge long before Edmonds proposed an algorithm for the non-bipartite case [3]. Regarding hyper-arc consistency, the problem has been resolved only for the bipartite case [1].

In this paper, we analyze the potential of using Emergent Self-Organizing Maps (ESOMs) based on Kohonen Self –Organizing maps in order to detect intrusive behaviours. The proposed approach combines machine learning and information visualization techniques to analyze network traffic and is based on classifying “normal” versus “abnormal” traffic. The results are promising as they show the ability of eSOMs to classify normal against abnormal behaviour regarding false alarms and detection probabilities.

This paper presents a conceptual mapping framework between a formal and visual process modelling language, Fundamental Business Process Modelling Language (FBPML), and the Web Services Ontology (OWL-S), aiming to bridge the gap between Enterprise Modelling methods and Semantic Web services. The framework is divided into a data model and a process model component. An implementation and an evaluation of the process model mapping are demonstrated.

Color plays an important role in web site design. The selection of effective chromatic combinations and the relation of color to the perceived aesthetic and emotional value of a web site is the focus of this paper. The subject of the reported research has been to define a model through which to be able to associate color combinations with specific desirable emotional and aesthetic values. The presented approach involves application of machine learning techniques on a rich data set collected during a number of empirical studies.

Fuzzy Cognitive Maps (FCMs) is a graphical model for causal knowledge representation. FCMs consist of nodes-concepts and weighted edges that connect the concepts and represent the cause and effect relationships among them. FCMs are used in complex problems involving causal relationships, which often include feedback, and where qualitative rather than quantitative measures of influences are available. They have used for decision support to determine a final state given a qualitative initial knowledge for nodes and weighted edges. A first study on introducing Interval analysis in the FCM framework has been attempted and it is presented in this work. Here a new structure for FCM is proposed with interval weights and a new method for processing interval data input for FCMs is proposed.

E-Learning service providers produce or collect digital learning resources, derive metadata for their description, and reuse and organize them in repositories. This paper proposes a data mining approach to discover relationships between the learning resources metadata. In particular, it presents and evaluates methods for clustering learning resources and providing controlled vocabularies for each class description. The derived classes and vocabularies contribute to the semantic interoperability in learning resource interchanges.

This paper, based on the organizational model proposed in [2], investigates the organization oriented programming paradigm. The approach proposed, in contrast to other approaches, emphasizes on group thinking. To show how the organization oriented programming paradigm is applied the paper describes the implementation of the asynchronous backtracking algorithm used in distributed CSPs.

In this paper we present a multimodal Greek Sign Language (GSL) recognizer. The system can recognize either signs or finger-spelled words of GSL, forming sentences of GSL. A vocabulary of 54 finger-spelled words together with 17 signs, giving a total of 71 signs/words, is used. The system has been tested on various grammars and the recognition rates we achieved exceeded 89% in most cases.

State of the art artificial agents rely heavily on human intervention for performing vision-language integration; apart from being cost and effort effective, this intervention deprives artificial agents from the ability to react intelligently and to show intentionality when engaged in situated multimodal communication. In this paper, we suggest an alternative way of building vision-language integration prototypes with limited human intervention. The suggestions have emerged from the development of such a prototype for the verbalisation of visual scenes in a property-surveillance task.

In this short paper we suggest that Population P Systems and Communication X-machines may be combined into one hybrid formal method which facilitates the correct specification of reconfigurable multi-agent systems.

This paper proposes an intelligent combination of neural network theory and financial statistical models for the detection of arbitrage opportunities in a group of stocks. The proposed intelligent methodology is based on a class of neural network-GARCH autoregressive models for the effective handling of the dynamics related to the statistical mispricing between relative stock prices. The performance of the proposed intelligent trading system is properly measured with the aid of profit & loss diagrams.

A faceted taxonomy is a set of taxonomies each describing the application domain from a different (preferably orthogonal) point of view. CTCA is an algebra that allows specifying the set of meaningful compound terms (meaningful conjunctions of terms) over a faceted taxonomy in a flexible and efficient manner. However, taxonomy updates may turn a CTCA expression

e

ill-formed and may turn the compound terms specified by

e

to no longer reflect the domain knowledge originally expressed in

e

. This paper shows how we can revise

e

after a taxonomy update and reach an expression

e

′ that is both well-formed and whose semantics (compound terms defined) is as close as possible to the semantics of the original expression

e

before the update.

In spectral datasets, such as those consisting of MR spectral data derived from MS lesions, neighboring features tend to be highly correlated, suggesting the data lie on some low-dimensional space. Naturally, finding such low-dimensional space is of interest. Based on this real-life problem, this paper extracts an abstract problem, neighboring feature clustering (NFC). Noticeably different from traditional clustering schemes where the order of features doesn’t matter, NFC requires that a cluster consist of neighboring features, that is features that are adjacent in the original feature ordering. NFC is then reduced to a piece-wise linear approximation problem. We use minimum description length (MDL) method to solve this reduced problem. The algorithm we proposed works well on synthetic datasets. NFC is an abstract problem. With minor changes, it can be applied to other fields where the problem of finding piece-wise neighboring groupings in a set of unlabeled data arises.