Artificial Intelligence: Methodology, Systems, and Applications

In this research, we investigate the issue of efficient detection of similar academic papers. Given a specific paper, and a corpus of academic papers, most of the papers from the corpus are filtered out using a fast filter method. Then, 47 methods (baseline methods and combinations of them) are applied to detect similar papers, where 34 of the methods are variants of new methods. These 34 methods are divided into three new method sets: rare words, combinations of at least two methods, and compare methods between portions of the papers. Results achieved by some of the 34 heuristic methods are better than the results of previous heuristic methods, comparing to the results of the “Full Fingerprint” (FF) method, an expensive method that served as an expert. Nevertheless, the run time of the new methods is much more efficient than the run time of the FF method. The most interesting finding is a method called CWA(1) that computes the frequency of rare words that appear only once in both compared papers. This method has been found as an efficient measure to check whether two papers are similar.

Derivational relations are an important part of the lexical semantics system in many languages, especially those of rich inflection. They represent wide variety of semantic oppositions. Analysis of morphological word forms in terms of prefixes and suffixes provides limited information about their semantics. We propose a method of semantic classification of the potential derivational pairs. The method is based on supervised learning, but requires only a list of word pairs assigned to the derivational relations. The classification was based on a combination of features describing distribution of a derivative and derivational base in a large corpus together with their morphological and morpho-syntactic properties. The method does not use patterns based on close co-occurrence of a derivative and its base. Two classification schemes were evaluated: a multiclass and a cascade of binary classifiers, both expressed good performance in experiments on the selected nominal derivational relations.

The paper presents an algorithm of the automated wordnet expansion which can utilise results produced by both pattern-based and Distributional Similarity methods. It is based on the assumption that all relation extraction methods express some error, so we cannot identify the exact place (synset) for a new lemma on their bases, but an area (a wordnet subgraph). Support for a particular attachment point generated by knowledge should be expanded on the surrounding synsets. Moreover, the wordnet structure is modelled on the level of links between lexical units. Evaluation of the algorithm and comparison with top algorithm from literature in large scale experiments on Princeton WordNet is presented.

Choosing model parameters is an important issue for solving real word problems. Wrong parameter values result in low performance of employed model. Usually, parameters are chosen manual, but one can employ metaheuristics for searching the parameter space in more systematic and automated way. In this paper we test a few optimisation methods such as Evolutionary Algorithms, Tabu Search, Hill Climbing and Simulated Annealing for setting parameters of models in two problems in the domain of Natural Language Processing. Metaheuristics used significantly improve performance in comparison to the default parameter selected manually by domain experts.

Having a very large volume of unstructured text documents representing different opinions without knowing which document belongs to a certain category, clustering can help reveal the classes. The presented research dealt with almost two millions of opinions concerning customers’ (dis)satisfaction with hotel services all over the world. The experiments investigated the automatic building of clusters representing positive and negative opinions. For the given high-dimensional sparse data, the aim was to find a clustering algorithm with a set of its best parameters, similarity and clustering-criterion function, word representation, and the role of stemming. As the given data had the information of belonging to the positive or negative class at its disposal, it was possible to verify the efficiency of various algorithms and parameters. From the entropy viewpoint, the best results were obtained with

k

-means using the binary representation with the cosine similarity,

idf

, and

H2

criterion function, while stemming played no role.

This paper presents a research prototype for temporal event information extraction from hospital discharge letters in Bulgarian. An algorithm for extraction of primitive events automatically sets markers for patients’ complaints, drug treatment and diagnoses with precision about 90%. Specific domain knowledge is further used to generate compound events and to identify some relations between event time sequences. Absolute and relative time information enables ordering the generated compound events using semi-intervals and fuzzy logic. Some negated events are analyzed as well to better structure the patient history.

Emotions are reflected both in verbal and written communication. If in the first case they can be easier to trace due to some specific features (body language, voice tone or inflections), in the second it can be quite tricky to grasp the underlying emotions carried by a written text. Therefore we propose a novel automatic method for analyzing emotions induced by texts, more specifically a reader’s most likely emotional state after reading a news article. In other words, our goal is to determine how reading a piece of news affects a person’s emotional state and to adjust these values based on his/her current state. From a more technical perspective, our system (

Emo2

–

Emo

tions

Mo

nitor) combines a context independent approach (actual evaluation of the news employing specific natural language processing techniques and Latent Semantic Analysis) with the influences of user’s present emotional state estimated through his/her specific feedback for building a more accurate image of a person’s emotional state.

In this research we present a configurable novel framework based on an enhanced heterogeneous hierarchical social fabric influence function embedded in Cultural Algorithms, as a powerful vehicle for the solution of complex problems. We motivate the discussion by investigating the extent to which these emergent phenomena are also visible within novel hybrid complex composition environments whose properties and complexity can be blurred and controlled easily, for the sake of overcoming any shortcomings of existing test functions that some of the current algorithms take advantage of during the optimization process. This environmental complexity induces an increase in the complexity of social roles within our system. We demonstrate how the well-configured hierarchical social fabric enhances Cultural Algorithms performance relative to other evolutionary algorithms from the literature.

This paper presents a novel peer-to-peer architecture for decentralized Online Social Network and a mechanism that allows each node to filter out irrelevant social data, while ensuring a level of serendipity, by letting important information pass, even if it does not fall in the areas of interest of the user. The evaluation of the approach, using an Erlang simulation with 2318 nodes shows that it works as it was designed to: with the increasing number of social data passing through the network, the nodes learn to filter out irrelevant data, while serendipitous important data is able to pass through the network. Future work will implement the mechanism in a decentralized OSN and evaluate it in a real community of users, in an enterprise context.

We would like to introduce BEECON, an information and event extraction system for business intelligence. This is the first ontology-based system for business documents analysis that is able to detect 41 different types of business events from unstructured sources of information. The described system is intended to enhance business intelligence efficiency by automatically extracting relevant content such as business entities and events. In order to achieve it, we use natural language processing techniques, pattern recognition algorithms and hand-written detection rules. In our test set consisting of 190 documents with 550 events, the system achieved 95% precision and 67% recall in detecting all supported business event types from newspaper texts.

Methods for automated classification of chemical data depend on identifying interesting parts and properties. However, classes of chemical entities which are highly symmetrical and contain large numbers of homogeneous parts (such as carbon atoms) are not straightforwardly classified in this fashion. One such class of molecules is the fullerene family, which shows potential for many novel applications including in biomedicine. The Web Ontology Language

OWL

cannot be used to represent the structure of fullerenes, as their structure is not tree-shaped. While individual members of the fullerene class can be modelled in standard

FOL

, expressing the properties of the class as a whole (independent of the count of atoms of the members) requires second-order quantification. Given the size of chemical ontologies such as ChEBI, using second-order expressivity in the general case is prohibitively expensive to practical applications. To address these conflicting requirements, we introduce a novel framework in which we heterogeneously integrate standard ontological modelling with monadic second-order reasoning over chemical graphs, enabling various kinds of information flow between the distinct representational layers.

Recently, the approaches that combine semantic web ontologies and web 2.0 technologies have constituted a significant research field. We present in this paper an original approach concerning a technology that has recognized a great popularity in these recent years, we talk about folksonomies. Our aim in this contribution is propose new technique for the Social Semantic Web technologies in order to see how we can overcome the problem of tags’ ambiguity automatically in folksonomies even when we choose representing these latter with ontologies. We’ll also illustrate how we can enrich any folksonomy by a set of pertinent data to improve and facilitate the resources’ retrieval in these systems; all this with tackling another problem, we speak about spelling variations.

A cognitive map represent concepts linked by influences. This paper introduces a taxonomic cognitive map model that allows to define a cognitive map on a taxonomy. This taxonomy is used to provide simplified views of a map to help a user to understand it. The model we present also provides a mechanism to help a group of people to build a map. Our proposal consists of a method for building a synthesized map from individual cognitive maps provided by each participant, called designer, by exploiting the taxonomy and preferences over these designers.

Methods for automatic functional description of gene groups are useful tools supporting the interpretation of biological experiments. The RuleGO algorithm provides functional interpretation of gene groups in a form of logical rules including combinations of Gene Ontology terms in their premises. The number of rules generated by the algorithm is usually huge and additional methods of rule quality evaluation and filtration are required in order to select the most interesting ones. In the paper, we apply the multicriteria decision making UTA method to obtain a ranking of rules based on subjective expert opinion which is provided in a form of an ordered list of several rules. The presented approach is applied to the well known data set from microarray experiment and the results are compared with the standard RuleGO compound rule quality measure.

We propose a method for generating and learning agent controllers, which combines techniques from automated planning and reinforcement learning. An incomplete description of the domain is first used to generate a non-deterministic automaton able to act (sub-optimally) in the given environment. Such a controller is then refined through experience, by learning choices at non-deterministic points. On the one hand, the incompleteness of the model, which would make a pure-planning approach ineffective, is overcome through learning. On the other hand, the portion of the domain available drives the learning process, that otherwise would be excessively expensive. Our method allows to adapt the behavior of a given planner to the environment, facing the unavoidable discrepancies between the model and the environment. We provide quantitative experiments with a simulator of a mobile robot to assess the performance of the proposed method.

Linked Open Data movement is maturing. LOD cloud increases by billions of triples yearly. Technologies and guidelines about how to produce LOD fast, how to assure their quality, and how to provide vertical oriented data services are being developed (LOD2, LATC, baseKB). Little is said however about how to include reasoning in the LOD framework, and about how to cope with its diversity. This paper deals with this topic. It presents a data service – FactForge – the biggest body of general knowledge from LOD on which inference is performed. It has close to 16B triples available for querying, derived from about 2B explicit triples, after inference and some OWLIM repository specific optimization. We discuss the impacts of the reference layer of FactForge and inference on the diversity of the web of data, and argue for a new paradigm of data services based on linked data verticals, and on inferred knowledge.

We study in this paper the problem of global consistency for qualitative constraints networks (

QCN

s) of the Point Algebra (

PA

) and the Interval Algebra (

IA

). In particular, we consider the subclass

$\mathcal{S}_{\sf PA}$

corresponding to the set of relations of

PA

except the relations { < , = } and { > , = }, and the subclass

$\mathcal{S}_{\sf IA}$

corresponding to pointizable relations of

IA

one can express by means of relations of

$\mathcal{S}_{\sf PA}$

. We prove that path-consistency implies global consistency for

QCN

s defined on these subclasses. Moreover, we show that with the subclasses corresponding to convex relations, there are unique greatest subclasses of

PA

and

IA

containing singleton relations satisfying this property.

This paper presents the algorithm for induction of unordered sets of regression rules. It uses sequential covering strategy and dynamic reduction to classification approach. The main focus is put on quality measures which control the process of rule induction. We examined the effectiveness of nine quality measures. Moreover, we propose and compare three schemes of the prediction of target attribute value of examples covered by more than one rule. We also show rule filtration algorithm for the reduction of the number of generated rules. All experiments were carried out on 35 benchmark datasets.

We consider particle swarm optimization algorithm with a constriction coefficient and investigate its competence in finding matrix eigenvalues. We propose a new objective function formulation and several slight algorithm modifications to compute eigenvalues effectively. We compare obtained results with standard numerical procedures for eigenvalues calculation and illustrate presented critical analysis with numerical examples. The problem generates interesting and challenging objective function landscapes, so it is a good benchmark to test abilities of PSO variants.

The parametric uniform crossover is a general form of the uniform crossover operator by which the swapping probability of each locus could be controlled. The swap probability could control the amount of disruption of high order hyper planes. Several variations of selecto-recombinative genetic algorithms are proposed for controlling the swap probability in the parametric uniform crossover operator. The suitability of the operator for diversifying the population while reducing disruption of the good partial solutions is studied. The experiments showed significant improvement in the performance of the algorithms when the parametric uniform crossover were used in comparison to algorithms that uniform crossover have been used in them.

Inductive Logic Programming (ILP) deals with the problem of finding a hypothesis covering positive examples and excluding negative examples. It uses first-order logic as a uniform representation for examples and hypothesis. In this paper we propose a method to boost any ILP learning algorithm by first decomposing the set of examples to subsets and applying the learning algorithm to each subset separately, second, merging the hypotheses obtained for the subsets to get a single hypothesis for the complete set of examples, and finally refining this single hypothesis to make it shorter. The proposed technique significantly outperforms existing approaches.

Learning collaborative behaviors is an essential part of multi agent systems. One of the suitable techniques for learning collaborative behaviors is observational learning. This paper describes a hybrid method for learning teamwork behaviors from an expert team by observation. More specifically, this paper describes a technique based on implicit knowledge acquired from observational learning and some domain expertise knowledge. In our method an expert team is observed by a team of learners to learn plans and save them in a plan base. Learners then use case-based planning to effectively act based on learned plans in order to imitate the expert team. To evaluate our method a simulated soccer team is developed. We argue that this approach provides a powerful complement to existing teamwork learning methods, specifically in learning complex goal oriented behaviors.

This paper addresses a problem of finding portfolios that perform better than investment funds while showing similar behaviour. The quality of investment portfolio can be measured using various criteria such as the return and some kind of risk measurement. Investors seek to maximize return while minimizing risk. In order to achieve this goal various instruments are considered. One of the possibilities is to entrust the assets to an investment fund. Investment funds build their own portfolios of stocks, bonds, commodities, currencies, etc.

In this paper we consider the problem of finding a portfolio which outperforms a given investment fund with respect to both the return and the risk and which also behaves in a similar way to the given fund. The rationale behind such an approach is that investment strategies of mutual funds are prepared by experts and are therefore expected to be reasonably good in terms of both the return and the risk. To achieve the presented goal we use a multiobjective evolutionary algorithm with a dedicated ”division mutation” operator and a local search procedure. Presented method seems capable of building portfolios with desired qualities.

A new Modular Recurrent Trainable Neural Network (MRTNN) has been used for system identification of two-mass-resort-damper nonlinear oscillatory plant. The first MRTNN module identified the exponential part of the unknown plant and the second one - the oscillatory part of the plant. The plant has been controlled by a direct adaptive neural control system with integral term. The RTNN controller used the estimated parameters and states to suppress the plant oscillations and the static plant output control error is reduced by an I-term added to the control.

The intuitive mode of structuring melodies by humans is very hard to reproduce in the context of an automated method. The human brain can differentiate between the pitch, timbre and the attack of a musical note even if the listener doesn’t have prior knowledge of musical theory; the successions of these notes could easily be the base for recognizing various sections of a song. This paper tries to give some insight in the problem of automatic structuring of musical content, by applying some techniques of machine learning. The experiment followed a TOP-DOWN approach by applying the algorithms on 7 different genres, afterwards on one album of a particular genre and in the end on a single audio file of the same genre. After the automatic structure analysis was performed the accuracy of the results was tested by a performance evaluator and by a human component.

In the last few years, the multiagent system’s paradigm has been more and more used in various fields, specially in the simulation field (MAS). Whenever a new application came into being and has been validated by its review board, specialists usually want to reuse it, fully or partially, in order to cut down the time and price of developing similar application.

But this reuse is not as simple as expected. In a previous article, we proposed the DOM modeling to tackle modeling difficulties which arise in a complex system. However this solution has its limits as we will develop here. In this paper, we define a more complete agent modeling, based on the MVC design pattern, in order to to push back these limits.

Sparse, informative feature representation has become an extremely successful method in action feature detection. Such features make the task more manageable while providing increased robustness to noise and pose variation. As the feature points detected are numerous, thus affect the computational efficiency. In this paper we present an improvement of this idea by decreasing the number of key points. And then we combine the use of 3D SIFT and pLSA in action categories. To test the validation of our method, we show the framework we devise in detail, also present some behavior recognition results on the KTH dataset including boxing, handclapping, hand waving, walking, jogging, and running.

This paper presents a reactive multi-agent system for echelon platoon organization. Platoon systems are sets of vehicles that move together while keeping a predefined geometrical configuration without any material coupling. Each vehicle represents an autonomous agent that behaves based only on its own perceptions. The distributed platoon organization problem consists in defining the algorithms to be executed by each vehicle’s embedded software, in order to maintain the desired platoon configuration during displacements. Platoon systems found in literature deal generally with column formations adapted to urban or highway transportation systems. Other formations such as line, echelon,... can be encountered in fields like agriculture and the military. In this paper, we focus on the platoon echelon formation.

An approach based on a virtual vehicle-to-vehicle interaction model composed of two springs is proposed. Those virtual springs attach a platoon’s vehicle to its local leader, another platoon’s vehicle. In this work, five different spring’s attachment points are evaluated, to compare them and conclude about the more suitable one, depending on platoon’s trajectory. Eventually, from the evaluation results, it can be conceived to make attachment points evolve during platoon operation.

This paper discusses the problem of territory division in Hide-and-Seek games. To obtain an efficient seeking performance for multiple seekers, the seekers should agree on searching their own territories and learn to visit good hiding places first so that the expected time to find the hider is minimized. We propose a learning model using Reinforcement Learning in a hierarchical learning structure. Elemental tasks of planning the path to each hiding place are learnt in the lower layer, and then the composite task of finding the optimal sequence is learnt in the higher layer. The proposed approach is examined on a set of different maps and resulted in convergece to the optimal solution.

Human activities usually have a motive and are driven by goal directed sequence of actions. Recognizing and supporting human activities is an important challenge for ambient assisted living of elderly in their home environment. By understanding an activity as a sequence of actions, we explore action specification languages for recognizing human activities. In this setting, we analyze the role of the situative space model for modeling indoor human activities in terms of an action specification language.

In the context of road urban traffic management, the problem of parking spots search is a major issue because of its serious economic and ecological fallout. In this paper, we propose a multi-agent system that aims to decrease, for private vehicles drivers, the parking spots search time. In the system that we propose, a community of drivers shares information about spots availability. Our solution has been tested following different configurations. The first results show a decrease in parking spots search time.

Smart home is a sub branch of ambient intelligence technology. It was initially used to control environmental systems such as lighting and heating; but recently the use of smart technology has been strengthened and expanded such that the resulting technologies promise to revolutionarize daily human life by making people’s surroundings flexible and adaptive. This paper proposes a technique that uses the paradigm of nets within nets to model the behavior of a mobile robot which is sensitive, adaptive and responsive to the presence of a human person and which is able to provide one or more homework.

The paper describes a new method of constructing semantic expansions of search requests about the achievements and failures of active systems (organizations, people) for improving the results of Web search. This method is based on the theory of K-representations (knowledge representations), proposed by V.A. Fomichov - a new theory of designing semantic-syntactic analysers of natural language texts with the broad use of formal means for representing input, intermediary, and output data. The method uses an original formal model of a goals base – a knowledge base containing the information about the goals of active systems. The stated approach is implemented with the help of the Web programming language Java: an experimental search system AOS (Aspect Oriented Search) has been developed and tested.

This paper presents a massively parallel evolutionary algorithm with local search mechanism dedicated to dynamic optimization. Its application for solving Dynamic Traveling Salesman Problem (DTSP) is discussed. The algorithm is designed for many-core graphics processors with the Compute Unified Device Architecture (CUDA), which is a parallel computing architecture for nVidia graphics processors. Experiments on a number of benchmark DTSP problems confirmed the efficiency of the algorithm and the parallel computing model designed.

The paper considers automatic vision based modelling and recognition of sign language expressions using smaller units than words. Modelling gestures with subunits is similar to modelling speech by means of phonemes. To define the subunits a data–driven procedure is proposed. The procedure consists in partitioning time series of feature vectors obtained from video material into subsequences which form homogeneous clusters. The cut points are determined by an optimisation procedure based on quality assessment of the resulting clusters. Then subunits are selected in two ways: as clusters’ representatives or as hidden Markov models of clusters. These two approaches result in differences in classifier design. Details of the solution and results of experiments on a database of 101 Polish words and 35 sentences used at the doctor’s and in the post office are given. Our subunit–based classifiers outperform their whole–word–based counterpart, which is particularly evident when new expressions are recognised on the basis of a small number of examples.

In order to develop truly intelligent systems, it is necessary to improve their ability to understand non-verbal communication. We propose a novel framework to recognize individuals and emotions from gait, in order to improve HRI. We collected the motion data of the torso from 4 professional actors’ gait, using motion capture system, and 7 non-actors’ using 2 IMU sensors. We developed Feature Value Method which is a PCA based classifier and finally we achieved high recognition rate through cross-validation.