One of the major stumbling blocks for artificial intelligence remains the commonsense knowledge problem. It is not clear how we could go about building a program which has all the commonsense knowledge of the average human adult. This has led to growing interest in the ‘developmental’ approach, which takes its inspiration from nature (especially the human infant) and attempts to build a program which could develop its own knowledge and abilities through interaction with the world. The challenge here is to find a learning program which can continuously build on what it knows, to reach increasingly sophisticated levels of knowledge. This survey reviews work in this area, with the emphasis on those that focus on early learning, for example, sensorimotor learning. The concluding discussion assesses the progress thus far and outlines some key problems which have yet to be addressed, and whose solution is essential to achieve the goals of the developmental approach.

This article presents and motivates an extended ontology knowledge model which represents explicitly semantic information about concepts. This knowledge model results from enriching the standard conceptual model with semantic information which precisely characterises the concept's properties and expected ambiguities, including which properties are prototypical of a concept and which are exceptional, the behaviour of properties over time and the degree of applicability of properties to subconcepts. This enriched conceptual model permits a precise characterisation of what is represented by class membership mechanisms and helps knowledge engineers to determine, in a straightforward manner, the meta-properties holding for a concept. Meta-properties are recognised to be the main tool for a formal ontological analysis that allows us to build ontologies with a clean and untangled taxonomic structure.

This enriched semantics can prove useful to describe what is known by agents in a multi-agent system, and might facilitate the use of reasoning mechanisms on the knowledge that instantiates an ontology. These mechanisms can be used to solve ambiguities that can arise when agents with heterogeneous ontologies have to interoperate in order to perform a task.

This paper will summarize and analyze the work of the different research groups who have recently made significant contributions in using Reinforcement Learning techniques to learn dialogue strategies for Spoken Dialogue Systems (SDSs). This use of stochastic planning and learning has become an important research area in the past 10 years, since it promises automatic data-driven optimization of the behavior of SDSs that were previously hand-coded by expert developers. We survey the most important developments in the field, compare and contrast the different approaches, and describe current open problems.

In the last two decades various intelligent robotics competitions have become very popular. Arguably the most well-known of these are the robotic soccer competitions. In addition to their value in attracting media and capturing the minds of the general public, these competitions also provide benchmark problems for various robotics and artificial intelligence (AI) technologies. As with any benchmark, care must be taken that the benchmark does not introduce unwarranted biases. This paper critically evaluates the AI contributions made by various robotic competitions on AI research.

Recent advancements in location-based recommendation system (LBRS) and the availability of online applications, such as Twitter, Instagram, Foursquare, Path, and Facebook have introduced new research challenges in the area of LBRS. Use of content, such as geo-tagged media, point location-based, and trajectory-based information help in connecting the gap between the online social networking services and the physical world. In this article, we present a systematic review of the scientific literature of LBRS and summarize the efforts and contributions proposed in the literature. We have performed a qualitative comparison of the existing techniques used in the area of LBRS. We present the basic filtration techniques used in LBRS followed by a discussion on the services and the location features the LBRS utilizes to perform the recommendations. The classification of criteria for recommendations and evaluation metrics are also presented. We have critically investigated the techniques proposed in the literature for LBRS and extracted the challenges and promising research topics for future work.

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.

This article surveys existing practical implementations of both defeasible and argumentation-based reasoning engines and associated literature. We aim to summarize the current state of the art in the research area, show that there are many similiarities and connections between the various implementations and also highlight the differences regarding evaluation goals and strategies. An important goal of this paper is to argue for the need for well-designed empirical evaluations, as well as formal complexity analysis, in order to justify the practical applicability of a reasoning engine. There are indeed many challenges to be faced in developing implementations of argumentation. Not least of these is the inherent computational complexity of the formal models. We cover some of the ways these challenges have been addressed, and provide pointers for future directions in realizing the goal of practical argumentation.

This paper gives a state-of-the-art overview of the rapidly expanding field of user modelling in artificially intelligent systems. After showing how a user modelling component can improve parts of the processing of, and the interaction with, a large number of systems, the current situation in this field is sketched by means of a number of short descriptions of systems employing user modelling techniques. Next, a synthesis of this review is made by discussing the aspects on which the existing methods differ. It turns out that these aspects can best be approached from two points of view: the technical level and a more abstract, functional level. This dichotomy results in eight dimensions on which to compare the modelling methods. The synthesis is then completed by describing the existing modelling techniques, and classifying the reviewed systems. In the final section, current trends in the field are outlined and prerequisites for acceptance of user modelling on a larger scale are discussed.

Since the end of the 1990s, there has been an increasing interest in the application of artificial intelligence (AI) planning techniques to solve real-life problems. In addition to characteristics of academic problems, such as the need to reason about actions, real-life problems require detailed knowledge elicitation, engineering, and management. A systematic design process in which Knowledge and Requirements Engineering tools play a fundamental role is necessary in such applications. One of the main challenges in such design process, and consequently in the study of Knowledge Engineering in AI planning, has been the analysis of requirements and their subsequent transformation into an input-ready model for planners. itSIMPLE is a research project dedicated to the study of a project process to support the design phases of real-life planning models. In this paper, we give an overview of itSIMPLE focusing on the main translation processes among a minimal set of representations: from requirements represented in Unified Modeling Language (UML) to Petri Nets and from UML models to planning domain definition language for problem solving.

Logic-based Benders decomposition (LBBD) has improved the state of the art for solving a variety of planning and scheduling problems, in part by combining the complementary strengths of constraint programming and mixed integer programming (MIP). We undertake a computational analysis of specific factors that contribute to the success of LBBD, to provide guidance for future implementations. We study a problem class that assign tasks to multiple resources and poses a cumulative scheduling problem on each resource. We find that LBBD is at least 1000 times faster than state-of-the-art MIP on larger instances, despite recent advances in the latter. Further, we conclude that LBBD is most effective when the planning and scheduling aspects of the problem are roughly balanced in difficulty. The most effective device for improving LBBD is the inclusion of a subproblem relaxation in the master problem. The strengthening of Benders cuts also plays an important role when the master and subproblem complexity are properly balanced. These findings suggest future research directions.

Complex systems often exhibit emergent behaviour, unexpected macro-level behaviour caused by the interaction of micro-level components. In multiagent systems, these micro-level components may be autonomous agents and the emergent behaviour may be expressed as norms—patterns of behaviour that arise among the agents in response to their environment and each other. These emergent norms may be beneficial (e.g. by encouraging cooperative behaviour), or detrimental, but in either case it is useful to recognize these norms as they emerge and either encourage or discourage their establishment. We term this process engineering the emergence of norms and have identified three steps: the identification of a possible norm, evaluation of its benefit and its encouragement (or discouragement). This paper is an attempt to provide a survey of existing research related to these steps. We also provide an analysis of the approaches based upon their suitability for a variety of normative systems: we examine the requirements for agents to have autonomy over their choice of norms, the degree of observability required in the system, and the norm enforcement methods. The paper concludes with an discussion of open issues.

This paper presents a graph-based knowledge representation and reasoning language. This language benefits from an important syntactic operation, which is called a graph homomorphism. This operation is sound and complete with respect to logical deduction. Hence, it is possible to do logical reasoning without using the language of logic but only graphical, thus visual, notions. This paper presents the main knowledge constructs of this language, elementary graph-based reasoning mechanisms, as well as the graph homomorphism, which encompasses all these elementary transformations in one global step. We put our work in context by presenting a concrete semantic annotation application example.

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.

In qualitative reasoning research, much effort has been spent on developing representation and reasoning formalisms. Only recently, the process of constructing models in terms of these formalisms has been recognised as an important research topic of its own. Approaches addressing this topic are examined in this review. For this purpose a general model of the task of constructing qualitative models is developed that serves as a frame of reference in considering these approaches. Two categories of approaches are identified: model composition and model induction approaches. The former compose a model from predefined partial models and the latter infer a model from behavioural data. Similarities and differences between the approaches are discussed using the general task model as a reference. It appears that the majority of approaches focus on automating model construction entirely. Assessing and debugging a model in cooperation with a modeller is identified as an important topic for future research

Recent approaches to providing advisory knowledge-based systems with explanation capabilities are reviewed. The importance of explaining a system's behaviour and conclusions was recognized early in the development of expert systems. Initial approaches were based on the presentation of an edited proof trace to the user, but while helpful for debugging knowledge bases, these explanations are of limited value to most users. Current work aims to expand the kinds of explanation which can be offered and to embed explanations into a dialogue so that the topic of the explanation can be negotiated between the user and the system. This raises issues of mutual knowledge and dialogue control which are discussed in the review.

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.

Distributed Artificial Intelligence has been loosely defined in terms of computation by distributed, intelligent agents. Although a variety of projects employing widely ranging methodologies have been reported, work in the field has matured enough to reveal some consensus about its main characteristics and principles. A number of prominent projects are described in detail, and two general frameworks, the System conceptual model and the agent conceptual model, are used to compare the different approaches. The paper concludes by reviewing approaches to formalizing some of the more critical capabilities required by multi-agent interaction.

This paper explores the hypothesis that ontologies can be used to improve the capabilities and performance of on-board route planning for autonomous vehicles. We name a variety of general benefits that ontologies may provide, and list numerous specific ways that ontologies may be used in different components of our chosen infrastructure: the 4D/RCS system architecture developed at NIST. Our initial focus is on simple roadway driving scenarios where the controlled vehicle encounters objects in its path. Our approach is to develop an ontology of objects in the environment, in conjunction with rules for estimating the damage that would be incurred by collisions with the different objects in different situations. Automated reasoning is used to estimate collision damage; this information is fed to the route planner to help it decide whether to avoid the object. We describe our current experiments and plans for future work.

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.