Applying intelligent agent technologies to support human learning activities has been the subject of recent work that reaches across computer science and education disciplines. This article discusses agent-based approaches that have been designed to address a range of pedagogical and/or curricular tasks. Three types of agents are identified in the literature: pedagogical agents, peer-learning agents, and demonstrating agents. Features of each type are considered, as well as the systems in which these agents are incorporated, examining common and divergent goals, system and agent architectures, and evaluation methodologies. Open issues are highlighted, and future directions for this burgeoning interdisciplinary field are suggested.

Multi-instance (MI) learning is a variant of inductive machine learning, where each learning example contains a bag of instances instead of a single feature vector. The term commonly refers to the supervised setting, where each bag is associated with a label. This type of representation is a natural fit for a number of real-world learning scenarios, including drug activity prediction and image classification, hence many MI learning algorithms have been proposed. Any MI learning method must relate instances to bag-level class labels, but many types of relationships between instances and class labels are possible. Although all early work in MI learning assumes a specific MI concept class known to be appropriate for a drug activity prediction domain; this ‘standard MI assumption’ is not guaranteed to hold in other domains. Much of the recent work in MI learning has concentrated on a relaxed view of the MI problem, where the standard MI assumption is dropped, and alternative assumptions are considered instead. However, often it is not clearly stated what particular assumption is used and how it relates to other assumptions that have been proposed. In this paper, we aim to clarify the use of alternative MI assumptions by reviewing the work done in this area.

Rough set theory (RST) has enjoyed an enormous amount of attention in recent years and has been applied to many real-world problems including data mining, pattern recognition, and intelligent control. Much research has recently been carried out in respect of both the development of the underlying theory and the application to new problem domains. This paper attempts to summarize the advances in RST, its extensions, and their applications. It also identifies important areas which require further investigation. Typical example application domains are examined which demonstrate the success of the application of RST to a wide variety of areas and disciplines, and which also exhibit the strengths and limitations of the respective underlying approaches.

Planning, scheduling and constraint satisfaction are important areas in artificial intelligence (AI) with broad practical applicability. Many real-world problems can be formulated as AI planning and scheduling (P&S) problems, where resources must be allocated to optimize overall performance objectives. Frequently, solving these problems requires an adequate mixture of planning, scheduling and resource allocation to competing goal activities over time in the presence of complex state-dependent constraints. Constraint satisfaction plays an important role in solving such real-life problems, and integrated techniques that manage P&S with constraint satisfaction are particularly useful. Knowledge engineering supports the solution of such problems by providing adequate modelling techniques and knowledge extraction techniques for improving the performance of planners and schedulers. Briefly speaking, knowledge engineering tools serve as a bridge between the real world and P&S systems.

The role concept represents a useful approach that has been exploited in different agent-based systems, in particular applied to interactions between agents. There are some requirements that are important for the development of agent-based applications using roles, for instance the support for the analysis, the design and the implementation phases. We have considered and compared different role-based proposals in the literature, and this paper presents a survey of the most spread ones. We explain each proposal and point out if and how it meets the identified requirements. Far from deciding the best proposal, our aim is to present the advantages and drawbacks of several proposals to designers and developers, so that they can make the best choice with regard to their needs.

In machine learning and its risk-sensitive applications (e.g. medicine, engineering, business), the reliability estimates for individual predictions provide more information about the individual prediction error (the difference between the true label and regression prediction) than the average accuracy of predictive model (e.g. relative mean squared error). Furthermore, they enable the users to distinguish between more and less reliable predictions. The empirical evaluations of the existing individual reliability estimates revealed that the successful estimates’ performance depends on the used regression model and on the particular problem domain. In the current paper, we focus on that problem as such and propose and empirically evaluate two approaches for automatic selection of the most appropriate estimate for a given domain and regression model: the internal cross-validation approach and the meta-learning approach. The testing results of both approaches demonstrated an advantage in the performance of dynamically chosen reliability estimates to the performance of the individual reliability estimates. The best results were achieved using the internal cross-validation procedure, where reliability estimates significantly positively correlated with the prediction error in 73% of experiments. In addition, the preliminary testing of the proposed methodology on a medical domain demonstrated the potential for its usage in practice.

Up to now, many data mining and knowledge discovery methodologies and process models have been developed, with varying degrees of success. In this paper, we describe the most used (in industrial and academic projects) and cited (in scientific literature) data mining and knowledge discovery methodologies and process models, providing an overview of its evolution along data mining and knowledge discovery history and setting down the state of the art in this topic. For every approach, we have provided a brief description of the proposed knowledge discovery in databases (KDD) process, discussing about special features, outstanding advantages and disadvantages of every approach. Apart from that, a global comparative of all presented data mining approaches is provided, focusing on the different steps and tasks in which every approach interprets the whole KDD process. As a result of the comparison, we propose a new data mining and knowledge discovery process named refined data mining process for developing any kind of data mining and knowledge discovery project. The refined data mining process is built on specific steps taken from analyzed approaches.

During recent years, the development of new techniques for constraint satisfaction, planning, and scheduling has received increased attention, and substantial effort has been invested in trying to exploit such techniques to find solutions to real-life problems. In this paper, we present a survey on constraint satisfaction, planning, and scheduling from the Artificial Intelligence point of view. In particular, we present the main definitions and techniques, and discuss possible ways of integrating such techniques. We also analyze the role of constraint satisfaction in planning and scheduling, and hint at some open research issues related to planning, scheduling, and constraint satisfaction.

There are many approaches for solving planning problems. Many of these approaches are based on ‘brute force’ search methods and they usually do not care about structures of plans previously computed in particular planning domains. By analyzing these structures, we can obtain useful knowledge that can help us find solutions to more complex planning problems. The method described in this paper is designed for gathering macro-operators by analyzing training plans. This sort of analysis is based on the investigation of action dependencies in training plans. Knowledge gained by our method can be passed directly to planning algorithms to improve their efficiency.

The knowledge discovery and data mining (KDDM) process models describe the various phases (e.g. business understanding, data understanding, data preparation, modeling, evaluation and deployment) of the KDDM process. They act as a roadmap for implementation of the KDDM process by presenting a list of tasks for executing the various phases. The checklist approach of describing the tasks is not adequately supported by appropriate tools, which specify ‘how’ the particular task can be implemented. This may result in tasks not being implemented. Another disadvantage is that the long checklist does not capture or leverage the dependencies that exist among the various tasks of the same and different phases. This not only makes the process cumbersome to implement, but also hinders possibilities for semi-automation of certain tasks. Given that each task in the process model serves an important goal and even affects the execution of related tasks due to the dependencies, these limitations are likely to negatively affect the efficiency and effectiveness of KDDM projects. This paper proposes an improved KDDM process model that overcomes these shortcomings by prescribing tools for supporting each task as well as identifying and leveraging dependencies among tasks for semi-automation of tasks, wherever possible.

Current classification problems that concern data sets of large and increasing size require scalable classification algorithms. In this study, we concentrate on several scalable, linear complexity classifiers that include one of the top 10 voted data mining methods, Naïve Bayes (NB), and several recently proposed semi-NB classifiers. These algorithms perform front-end discretization of the continuous features since by design they work only with nominal or discrete features. We address the lack of studies that investigate the benefits and drawbacks of discretization in the context of the subsequent classification. Our comprehensive empirical study considers 12 discretizers (two unsupervised and 10 supervised), seven classifiers (two classical NB and five semi-NB), and 16 data sets. We investigate the scalability of the discretizers and show that the fastest supervised discretizers fast class-attribute interdependency maximization (FCAIM), class-attribute interdependency maximization (CAIM), and information entropy maximization (IEM) provide discretization schemes with the highest overall quality. We show that discretization improves the classification accuracy when compared against the two classical methods, NB and Flexible Naïve Bayes (FNB), executed on the raw data. The choice of the discretization algorithm impacts the significance of the improvements. The MODL, FCAIM, and CAIM methods provide statistically significant improvements, while the IEM, Class-attribute contingency coefficient (CACC), and Khiops discretizers provide moderate improvements. The most accurate classification models are generated by the Averaged one-dependence estimators (AODEsr) classifier followed by AODE and HNB (Hidden Naïve Bayes). AODEsr run on data discretized with MODL, FCAIM, and CAIM provides statistically significantly better accuracies than both the classical NB methods. The worst results are obtained with the NB, FNB, and LBR (Lazy Bayes rule) classifiers. We show that although the time to build the discretization scheme could be longer than the time to train the classifier, the completion of the entire process (to discretize data, compute the classifier, and predict test instances) is often faster than the NB-based classification of the continuous instances. This is because the time to classify test instances is an important factor that is positively influenced by discretization. The biggest positive influence, both on the accuracy and the classification time, is associated with the MODL, FCAIM, and CAIM algorithms.

Web-based expert systems (WBESs) provide the benefits of both expert system technology and web technology. The use of web services to deliver functionalities of WBESs allows the integration of these systems in web-portals. WBESs are used in a diversity of areas like engineering, management, medicine, agriculture, education, tourism, finance etc. A study on the various features of WBESs like knowledge-representation, reasoning, languages, implementation tools, use of various web services-related processes such as discovery, selection, composition etc. can result into some interesting conclusions. Presented work tries to fulfill the same purpose. The paper presents various WBESs classified according to their use of domain. The comparisons, observations, and assessments of these systems are presented by emphasizing the above-mentioned features of WBESs. A discussion on different web services-related processes and some representative techniques for performing each has also been provided to clarify their use in the WBESs. On the basis of assessments and surveys from different perspectives, some remarkable conclusions are drawn.

To foster effective use of artificial intelligence planning and scheduling (P&S) systems in the real world, it is of great importance to both (a) broaden direct access to the technology for the end users and (b) significantly increase their trust in such technology. Automated P&S systems often bring solutions to the users that are neither ‘obvious’ nor immediately acceptable to them. This is because these tools directly reason on causal, temporal, and resource constraints; moreover, they employ resolution processes designed to optimize the solution with respect to non-trivial evaluation functions. Knowledge engineering environments aim at simplifying direct access to the technology for people other than the original system designers, while the integration of validation and verification (V&V) capabilities in such environments may potentially enhance the users’ trust in the technology. Somehow, V&V techniques may represent a complementary technology, with respect to P&S, that contributes to developing richer software environments to synthesize a new generation of robust problem-solving applications. The integration of V&V and P&S techniques in a knowledge engineering environment is the topic of this paper. In particular, it analyzes the use of state-of-the-art V&V technology to support knowledge engineering for a timeline-based planning system called MrSPOCK. The paper presents the application domain for which the automated solver has been developed, introduces the timeline-based planning ideas, and then describes the different possibilities to apply V&V to planning. Hence, it continues by describing the step of adding V&V functionalities around the specialized planner, MrSPOCK. New functionalities have been added to perform both model validation and plan verification. Lastly, a specific section describes the benefits as well as the performance of such functionalities.

Over the last two decades, qualitative reasoning (QR) has become an important domain in Artificial Intelligence. QDE (Qualitative Differential Equation) model learning (QML), as a branch of QR, has also received an increasing amount of attention; many systems have been proposed to solve various significant problems in this field. QML has been applied to a wide range of fields, including physics, biology and medical science. In this paper, we first identify the scope of this review by distinguishing QML from other QML systems, and then review all the noteworthy QML systems within this scope. The applications of QML in several application domains are also introduced briefly. Finally, the future directions of QML are explored from different perspectives.

With the rapid growth in the quantity and complexity of scientific knowledge available for scientists, and allied professionals, the problems associated with harnessing this knowledge are well recognized. Some of these problems are a result of the uncertainties and inconsistencies that arise in this knowledge. Other problems arise from heterogeneous and informal formats for this knowledge. To address these problems, developments in the application of knowledge representation and reasoning technologies can allow scientific knowledge to be captured in logic-based formalisms. Using such formalisms, we can undertake reasoning with the uncertainty and inconsistency to allow automated techniques to be used for querying and combining of scientific knowledge. Furthermore, by harnessing background knowledge, the querying and combining tasks can be carried out more intelligently. In this paper, we review some of the significant proposals for formalisms for representing and reasoning with scientific knowledge.

The CNT framework (Constraint Network on Timelines) has been designed to model discrete event dynamic systems and the properties one knows, one wants to verify, or one wants to enforce on them. In this article, after a reminder about the CNT framework, we show its modeling power and its ability to support various modeling styles, coming from the planning, scheduling, and constraint programming communities. We do that by producing and comparing various models of two mission management problems in the aerospace domain: management of a team of unmanned air vehicles and of an Earth observing satellite.