Advances in Artificial Intelligence - SBIA 2012

In this article, we propose a formal methodology for knowledge representation in DeLP, that defines a set of guidelines to be used during this phase. Our methodology results in an key tool to improve DeLP’s applicability to concrete domains.

Belief revision has been extensively studied in the last thirty years. While there are many results in the literature comparing different operators from a theoretical point of view, there is no study of how the different operators perform in practice. In this paper, we propose a framework for empirical testing of belief change operators. The idea is that any operator can be quickly implemented making use of the available API and then tested for multiple scenarios. We illustrate the use of the framework with a case study comparing partial meet and kernel contraction operators.

The autonomous cognitive development paradigm applied to cognitive agent’s design assumes that no prior knowledge should be embedded while the agent’s cognitive system is conceived. The use of this paradigm is an attempt to avoid the symbol grounding problem faced by some cognitive agent designs. We assume that the cognitive system of a sensorimotor agent is not aware of what it’s sensors sense or what it’s actuators do, or even where they are connected to in the agent’s body. Using these prerogatives we present a sensorimotor domain approach that can be used to build agent’s controls and to construct agents’ cognitive systems. This proposal is then compared to an approach for sensorimotor control system implementation, here called sensorimotor functional approach, under the perspective of the autonomous cognitive development paradigm. In order to highlight the advantages of our proposal and to show how it can bypass the limitations of the functional approach, two simple examples are presented.

Although there are increasingly more initiatives for the generation of semantic knowledge based on user participation, there is still a shortage of platforms for regular users to create applications on which semantic data can be exploited and generated automatically. We propose an architecture, called Semantic Maps (SeMaps), for assisting the authoring and hosting of applications in which the maps combine the aggregation of a Geographic Information System and crowd-generated content (called here crowd maps). In these systems, the digital map works as a blackboard for accommodating stories told by people about events they want to share with others typically participating in their social networks. SeMaps offers an environment for the creation and maintenance of sites based on crowd maps with the possibility for the user to characterize semantically that which s/he intends to mark on the map. The designer of a crowd map, by informing a linguistic expression that designates what has to be marked on the maps, is guided in a process that aims to associate a concept from a common-sense base to this linguistic expression. Thus, the crowd maps start to have dominion over common-sense inferential relations that define the meaning of the marker, and are able to make inferences about the network of linked data. This makes it possible to generate maps that have the power to perform inferences and access external sources (such as DBpedia) that constitute information that is useful and appropriate to the context of the map. In this paper we describe the architecture of SeMaps and how it was applied in a crowd map authoring tool.

Natural-language preference expressions, not yet exploited by existing preference reasoning approaches, match the way users express preferences in many scenarios and potentially improve automated decision making. Further, the preferences provided are often not sufficient to make a choice on behalf of users, as trade-offs are resolved with psychological processes employed in light of available options. We thus propose a decision making technique that reasons about preferences expressed in a user-centric language and incorporates principles of trade-off contrast and extremeness aversion, as in human decision-making.

During the last 5 years, research on Human Activity Recognition (HAR) has reported on systems showing good overall recognition performance. As a consequence, HAR has been considered as a potential technology for e-health systems. Here, we propose a machine learning based HAR classifier. We also provide a full experimental description that contains the HAR wearable devices setup and a public domain dataset comprising 165,633 samples. We consider 5 activity classes, gathered from 4 subjects wearing accelerometers mounted on their waist, left thigh, right arm, and right ankle. As basic input features to our classifier we use 12 attributes derived from a time window of 150ms. Finally, the classifier uses a committee AdaBoost that combines ten Decision Trees. The observed classifier accuracy is 99.4%.

Time series data is a very common kind of data in many different fields. In particular, unknown frequent pattern discovery is one of the core activities in many time series mining algorithms. Several solutions to pattern discovery have been proposed so far. However, all solutions assume centralized dataset. With increasingly development of network technology distributed data analysis has become popular, raising issues like scalability and cost minimization. Additionally, some scenarios such as mining distributed medical or financial data involves the question of how to preserve data privacy. In this paper, we present a density based pattern discovery algorithm for time series, which is shown to be efficient and privacy-preserving.

In multi-label learning, each example in the dataset is associated with a set of labels, and the task of the generated classifier is to predict the label set of unseen examples. Feature selection is an important task in machine learning, which aims to find a small number of features that describes the dataset as well as, or even better, than the original set of features does. This can be achieved by removing irrelevant and/or redundant features according to some importance criterion. Although effective feature selection methods to support classification for single-label data are abound, this is not the case for multi-label data. This work proposes two multi-label feature selection methods which use the filter approach. This approach evaluates statistics of the data independently of any particular classifier. To this end, ReliefF, a single-label feature selection method and an adaptation of the Information Gain measure for multi-label data are used to find the features that should be selected. Both methods were experimentally evaluated in ten benchmark datasets, taking into account the reduction in the number of features as well as the quality of the generated classifiers, showing promising results.

This work addresses the issue of automatic organic component detection and segmentation in confocal microscopy images. The proposed method performs cellular/parasitic identification through adaptive segmentation using a two-level Otsu’s Method. Segmented regions are divided using a rule-based classifier modeled on a decreasing harmonic function and a Support Vector Machine trained with features extracted from several Gaussian mixture models of the segmented regions. Results indicate the proposed method is able to count cells and parasites with accuracies above 90%, as well as perform individual cell/parasite detection in multiple nucleic regions with approximately 85% accuracy. Runtime measures indicate the proposed method is also adequate for real-time usage.

Cluster in graphs is densely connected group of vertices sparsely connected to other groups. Hence, for prediction of a future link between a pair of vertices, these vertices common neighbors may play different roles depending on if they belong or not to the same cluster. Based on that, we propose a new measure (WIC) for link prediction between a pair of vertices considering the sets of their intra-cluster or within-cluster (W) and between-cluster or inter-cluster (IC) common neighbors. Also, we propose a set of measures, referred to as W forms, using only the set given by the within-cluster common neighbors instead of using the set of all common neighbors as usually considered in the basic local similarity measures. Consequently, a previous clustering scheme must be applied on the graph. Using three different clustering algorithms, we compared WIC measure with ten basic local similarity measures and their counterpart W forms on ten real networks. Our analyses suggest that clustering information, no matter the clustering algorithm used, improves link prediction accuracy.

Recent innovations have enabled ever increasing amounts of data to be collected and stored, leading to the problem of extracting knowledge from it. Clustering techniques help organizing and understanding such data, and parallelization of such may reduce the cost of achieving this goal or improve on the result. This works presents the parallel implementation of the HACO clustering method, analyzing process of parallelization and its results with different topologies and communication strategies.

Incremental learning of topic hierarchies is very useful to organize and manage growing text collections, thereby summarizing the implicit knowledge from textual data. However, currently available methods have some limitations to perform the incremental learning phase. In particular, when the initial topic hierarchy is not suitable for modeling the data, new documents are inserted into inappropriate topics and this error gets propagated into future hierarchy updates, thus decreasing the quality of the knowledge extraction process. We introduce a method for obtaining more robust initial topic hierarchies by using consensus clustering. Experimental results on several text collections show that our method significantly reduces the degradation of the topic hierarchies during the incremental learning compared to a traditional method.

This paper proposes a new method for content-based image retrieval that uses a computational model of visual attention and genetic algorithm to find a given object in a set of images with different backgrounds. This method is composed by three main modules: a visual attention model that is quite robust against affine transformations; a color-based schematic representation of visual information; and a genetic algorithm that optimizes several parameters of the visual attention model in order to focus the attention mechanism on those regions of the image where it is most likely that a given object is present. The proposed method is validated through several experiments, and these experiments show that it can find the images that contain the sought object as well as the position and scale of the object in these images.

In recent years many studies have been proposed for knowledge discovery in time series. Most methods use some technique to transform raw data into another representation. Symbolic representations approaches have shown effectiveness in speedup processing and noise removal. The current most commonly used algorithm is the Symbolic Aggregate Approximation (SAX). However, SAX doesn’t preserve the slope information of the time series segments because it uses only the Piecewise Aggregate Approximation for dimensionality reduction. In this paper, we present a symbolic representation method to dimensionality reduction and discretization that preserves the behavior of slope characteristics of the time series segments. The proposed method was compared with the SAX algorithm using artificial and real datasets with 1-nearest-neighbor classification. Experimental results demonstrate the method effectiveness to reduce the error rates of time series classification and to keep the slope information in the symbolic representation.

In comparison to single agent learning, reinforcement learning in a multiagent scenario is more challenging, since there is an increase in the space of combination of actions that may have to be explored before agents learn an efficient policy. Among other approaches, there has been a proposition to address this problem by means of biasing the exploration. We follow this track using an organizational structure where low-level agents mainly use reinforcement learning, while also getting recommendations from agents possessing a broader view. These agents keep a base of cases in order to give such recommendations, orchestrating the process. We show that this approach is able to accelerate and improve learning in penalty games, a especial case of coordination games.

Social norms have become one of the most promising approaches that use an informal social control to ensure a desirable social order in open multi-agent systems. In these systems, autonomous and heterogeneous entities work towards similar or different goals. Norms regulate the behaviour of agents by defining obligations and prohibitions, and by creating rewards and penalties to encourage the agents to meet theses constraints. The development of autonomous normative agents, i. e., agents able to take decisions by following their motivations related to their goals and taking into account the system norms, has proven to be significantly more challenging than the design of traditional agents. In this paper, we introduce an architectural model that provides a set of functions to support the agent on the reasoning about the norms. These operations assist the agent to perceive the system’s norms, detect the fulfillment and violation of the norms while verifying their activation and deactivation, select the norms the agent intends to fulfill, identify and overcome conflicts among norms.

Most multi-agent systems engineering methodologies propose the clear definition of roles and organizations. However, in distributed environments where agents with distinct capabilities coexist and cooperate to solve problems, having a rigid organization structure makes the system less adaptable to changes and failures. Some of the approaches to deal with these difficulties include centralized coordination and planning and the use of homogeneous agent capabilities. These solutions oppose key benefits of multi-agent systems, especially the agents’ autonomy to interact and organize freely. In this paper, a novel approach is proposed where agents form and dissolve coalitions in a service-oriented environment while maintaining their autonomy. This allows the agent society to adjust to the demand for services and react to failures. To achieve this flexibility, a coalition formation mechanism for trust and reputation-aware multi-agent systems is employed. As agents interact, they establish a network of trusted peers that allows them to form stable coalitions with reduced risk of failures. Agents can also expand this network by exploring new partnerships based on the reputation of unknown agents that are recommended by these known peers. Experiments were performed to evaluate the proposal, with positive results in environments up to fifty agents under varying service demand and failure rates.

This work describes the study, analysis, modeling and implementation of information agents aiming to manage and share knowledge in a context of virtual Communities of Practice. The main goal is to define a decentralized platform to manage profiles of users and communities. Two types of agents have been modeled, one including the behaviors needed to agents representing communities and other including the behaviors needed to agents representing members of communities. These behaviors have access to the profiles, what allow agents to exchange information about interest and knowledge in order to receive recommendations according to similarities found. Some experiments are presented showing how the system works in a specific domain and the results obtained.

A promising approach to cope with the challenges that Word Sense Disambiguation brings is to use knowledge-intensive methods. Typically they rely on Wikipedia for supporting automatic concept identification. The exclusive use of Wikipedia as a knowledge base for word disambiguation and therefore the general identification of topics, however, have low accuracy vis-à-vis texts with diverse topics, as can be the case with blogs. This motivated us to propose a method for word disambiguation that, in addition to the use of Wikipedia, uses a common sense database. Use of this base enriches the definition of the concepts previously identified with the help of Wikipedia, and permits the definition of a similarity measure between concepts, which is characterized by verifying the similarity of two concepts from the viewpoint of conceptual proximity in the Wikipedia hierarchy, in addition to the proximity between such concepts in terms of the inferences that they can make. We show that by doing this, we improved the accuracy of automatic disambiguation of words compared with methods that do not use a common sense base.

Automatic Speech Recognition (ASR) is a complex task, which depends on language, vocabulary and context. In the navigation control of mobile robots, the set of possible interpretations for a command utterance may be reduced in favor of the recognition rate increase, if we consider that the robot’s work environment is quite defined and with constant elements. In this paper we propose a contextual model in addition to the acoustic and language models used by mainstream ASRs. We provide a whole mobile robot navigation system which use contextual information to improve the recognition rate of speech-based commands. Recognition accuracy has been evaluated by Word Information Lost (WIL) metric. Results show that the insertion of a contextual model provides a improvement around 3% on WIL.

A lot of scientists study the behavior of insect’s colony like ants, wasps and bees. Through these researches, it is possible to establish patterns used by a group of insects and apply these patterns in other domains. In this paper it will be showed the use of stigmergy in a rescue situation using the RoboCup Rescue simulator. We performed a set of experiments using a metaphor based on the behavior of an ant colony, where the communication between agents is done through the environment. We measured the performance of the ant-based algorithm, expecting to figure out the feasibility of using swarm intelligence in a rescue situation. We compared the results of using stigmergy against a multiagent system based on direct messages. The results showed that the use of stigmergy can outperform the use of direct messages.

The concept of optimization is present in several natural processes such as the evolution of species, the behavior of social groups and the ecological relationships of different animal populations. This work uses the concepts of habitats, ecological relationships and ecological successions to build a hybrid cooperative search algorithm, named ECO. The Artificial Bee Colony (ABC) and the Particle Swarm Optimization (PSO) algorithms were used in the experiments where benchmark mathematical functions were optimized. Results were compared with ABC and PSO running alone, and with both algorithms in a well known island model with ring topology, all running without the ecology concepts previously mentioned. The ECO algorithm performed better than the other approaches, especially as the dimensionality of the functions increase, possibly thanks to the ecological interactions (intra and inter-habitats) that enabled the co-evolution of populations. Results suggest that the ECO algorithm can be an interesting alternative for numerical optimization.

The combination of testing techniques is considered an effective strategy to evaluate the quality of a software product. However, the selection of which techniques to combine in a software project has been an interesting challenge in the software engineering field because the high number of techniques available at the technical literature. This paper presents an approach developed to support the combined selection of model-based testing techniques, applying multiobjective combinatorial optimization strategies, by determining the minimum dominating set in a bipartite and bi-weighted graph. Thus, an evolutionary strategy based on a multiobjective genetic algorithm is proposed to generate trade-off techniques subsets between the maximum coverage of software project characteristics and the minimum eventual effort to construct models used for test cases generation. In an empirical evaluation, our evolutionaryalgorithmstrategygavebetterresultsthanthepreviousapproaches.