We were faced with a shelved knowledge system at a health insurance company. The system used Bayesian statistics. For correct computations such a system is of course well suited, but not for the people who use it: people do not easily formulate their expertise in numerical probabilities. We addressed the question of whether it is possible to find out in which format experts do represent their knowledge, and if so, what format that is. This representation might then replace the format of numerical probabilities in the knowledge system's interface, increasing its ease of use. We found one method that includes, apart from a technique to elicit knowledge contents, techniques to also acquire its format: Cognitive Structure Analysis. In this contribution we discuss this method, use it in two studies, and show that it yields valid results. It gives insight into the format of experts' knowledge structures; experts turn out to use mainly frame representations.

For a number of years, researchers in AI and robotics have been sharing organisation principles for software development called “architectures” (Arkin, 1998; Hexmoor et al., 1997; Kortenkamp et al., 1998). Most recently, there has been an interest in extending the software architectures originally designed for robotic applications to accommodate autonomous operation beyond robotics (Pell et al., 1998). These architectures attempt to state principles for organising software for intelligent systems that exhibit cognitive functionalities such as natural language understanding, planning and learning, as well as organism-like functionalities to cope in the world such as reflexive and reactive behaviours.

Research on planning in AI can be separated into the two major areas: plan generation and plan representation. Most AI planners to date have been based on the STRIPS planning representation. This representation has a number of limitations. Much recent work in plan representation has addressed these limitations. It was shown that Decision Theory can be used to remove a number of the limitations. Furthermore, the decision theoretic framework provides a precise definition of rational behaviour. There remain open questions within decision theory regarding belief revision and causality. It should be noted that these problems are not artifacts of the representation. Rather they arise because the rich representation allows their formulation. Some work integrating AI and decision theoretic approaches to planning has been done but this remains a largely untouched research area.

We see two main avenues for fruitful research. First, the straightforward decision theoretic formulation of planning is computationally impractical. Techniques need to be developed to do efficient decision theoretic planning. Work in AI plan generation has exploited information contained the structure of qualitative representations to guide efficient plan construction. These techniques should be applied to decision theoretic representations as well. Second, AI has developed many representations that allow useful structuring of knowledge about the world. Decision Theory has concentrated on representing beliefs and desires. Integration of AI and decision theoretic representations would yield powerful representation languages. Some of the benefits of such work can already be seen in the research combining temporal and decision theoretic representations.

Building a probabilistic network for a real-life domain of application is a hard and time-consuming process, which is generally performed with the help of domain experts. As the scope and, hence, the size and complexity of networks are increasing, the need for proper management of the elicited domain knowledge becomes apparent. To study the usefulness of ontologies for this purpose, we constructed an ontology for the domain of oesophageal cancer, based on a real-life probabilistic network for the staging of cancer of the oesophagus and the knowledge elicited for its construction. In this paper, we describe the various components of our ontology and outline the benefits of using ontologies in engineering probabilistic networks.

The execution model of Prolog, the first popular language based on Horn Clauses, was designed for efficient evaluation on von Neumann architectures. An alternative process model of execution, better suited for parallel evaluation and reactive programming, has given rise to a new class of languages based on Horn Clause logic, concurrent logic languages. There appears to be a profusion of languages which claim to fall into this class and it is difficult for an initiate to appreciate why each is the way it is. One notable member of this class, FGHC, forms the cornerstone of the Japanese 5th Generation Initiative. Fortunately, the seemingly exponential growth in these languages is only an illusion. A finite number of synchronization mechanisms arise from attempting (or sometimes not attempting) to control two principle synchronization difficulties: the premature binding problem and the binding conflict problem. Suitable combinations of these synchronization mechanisms reproduce the languages of this family. A background knowledge of Prolog is assumed and some familiarity with the difficulties encountered in concurrency would be advantageous.

With the recent increase in the use of Unmanned Aerial Vehicles (UAVs) comes a surge of inexperienced aviators who may not have the requisite skills to react appropriately if weather conditions quickly change while their aircraft are in flight. This creates a dangerous situation, in which the pilot cannot safely land the vehicle. In this work we examine the use of the MAP-Elites algorithm to search for sets of weights for use in an artificial neural network. This neural network directly controls the thrust and pitching torque of a simulated 3-degree of freedom (2 linear, 1 rotational) fixed-wing UAV, with the goal of obtaining a smooth landing profile. We then examine the use of the same algorithm in high-wind conditions, with gusts up to 30 knots.

Our results show that MAP-Elites is an effective method for searching for control policies, and by evolving two separate controllers and switching which controller is active when the UAV is near-ground level, we can produce a wider variety of phenotypic behaviors. The best controllers achieved landing at a vertical speed of <1 m s−1 and at an angle of approach of <1° degree.

In this paper a number of alternative strategies for distributed/parallel association rule mining are investigated. The methods examined make use of a data structure, the T-tree, introduced previously by the authors as a structure for organizing sets of attributes for which support is being counted. We consider six different approaches, representing different ways of parallelizing the basic Apriori-T algorithm that we use. The methods focus on different mechanisms for partitioning the data between processes, and for reducing the message-passing overhead. Both ‘horizontal’ (data distribution) and ‘vertical’ (candidate distribution) partitioning strategies are considered, including a vertical partitioning algorithm (DATA-VP) which we have developed to exploit the structure of the T-tree. We present experimental results examining the performance of the methods in implementations using JavaSpaces. We conclude that in a JavaSpaces environment, candidate distribution strategies offer better performance than those that distribute the original dataset, because of the lower messaging overhead, and the DATA-VP algorithm produced results that are especially encouraging.

In this study, we designed a localization and obstacle avoidance system for humanoid robots in the Federation of International Robot-soccer Association (FIRA) HuroCup united soccer competition event. The localization is implemented by using grid points, gait, and steps to determine the positions of each robot. To increase the localization accuracy and eliminate the accumulated distance errors resulting from step counting, the localization is augmented with image pattern matching using a system model. The system also enables the robot to determine the ball’s position on the field using a color model of the ball. Moreover, to avoid obstacles, the robots calculate the obstacle distance using data extracted from real-time images and determine a suitable direction for movement. With the integration of this accurate self-localization algorithm, ball identification scheme, and obstacle avoidance system, the robot team is capable of accomplishing the necessary tasks for a FIRA soccer game.

It is now more than ten years since researchers in the US Knowledge Sharing Effort envisaged a future where complex systems could be built by combining knowledge and services from multiple knowledge bases and the first agent communication language, KQML, was proposed (Neches et al., 1991). This model of communication, based on speech acts, a declarative message content representation language and the use of explicit ontologies defining the domains of discourse (Genesereth & Ketchpel, 1994), has become widely recognised as having great benefits for the integration of disparate and distributed information sources to form an open, extensible and loosely coupled system. In particular, this idea has become a key tenet in the multi-agent systems research community.

The papers in this special issue represent an early view on different aspects of a new research field in computing science. They were presented as best papers in a series of European workshops called ‘Self-adaptive and autonomic systems’ (SAACS), which is co-located with the DEXA conference every year.

This paper presents trends and performance of opinion retrieval techniques proposed within the last 8 years. We identify major techniques in opinion retrieval and group them into four popular categories. We describe the state-of-the-art techniques for each category and emphasize on their performance and limitations. We then summarize with a performance comparison table for the techniques on different datasets. Finally, we highlight possible future research directions that can help solve existing challenges in opinion retrieval.

One challenge faced by reinforcement learning (RL) agents is that in many environments the reward signal is sparse, leading to slow improvement of the agent’s performance in early learning episodes. Potential-based reward shaping can help to resolve the aforementioned issue of sparse reward by incorporating an expert’s domain knowledge into the learning through a potential function. Past work on reinforcement learning from demonstration (RLfD) directly mapped (sub-optimal) human expert demonstration to a potential function, which can speed up RL. In this paper we propose an introspective RL agent that significantly further speeds up the learning. An introspective RL agent records its state–action decisions and experience during learning in a priority queue. Good quality decisions, according to a Monte Carlo estimation, will be kept in the queue, while poorer decisions will be rejected. The queue is then used as demonstration to speed up RL via reward shaping. A human expert’s demonstration can be used to initialize the priority queue before the learning process starts. Experimental validation in the 4-dimensional CartPole domain and the 27-dimensional Super Mario AI domain shows that our approach significantly outperforms non-introspective RL and state-of-the-art approaches in RLfD in both domains.

In this paper, we present an approach to solve the drawbacks of manual composition of software components. Our approach is applied within the jcolibri framework for building case-based reasoning (CBR) applications. We propose a system design process based on reusing templates obtained from previously designed CBR systems. Templates store the control flow of the CBR applications and include semantic annotations conceptualizing its behavior and expertise. We use CBR ontology to formalize syntactical, semantical and pragmatical aspects of the reusable components of the framework. The ontology vocabulary facilitates an annotation process of the components and allows to reason about their composition, facilitating the semi-automatic configuration of complex systems from their composing pieces.

Negotiation in a multiagent system is a topic of active interest for enabling the allocation of scarce resources among autonomous agents. This paper presents a discussion of the research on negotiation criteria, which puts in context the contributions to resource allocation from the fields of economics, mathematics, and multiagent systems. We group the criteria based on how they relate to each other as well as their historical origin. In addition, we present three new criteria: verifiability, dimensionality, and topology. The criteria are organized into five categories. The allocation category contains criteria concerning fairness and envy-freeness with respect to how resources are allocated to agents. The protocol category covers criteria for stability, strategy-proofness, and communication costs. The procedure category includes criteria about the complexity of allocation procedures. The resource category has criteria for the properties that various resources can take and how they affect allocation. The paper concludes with a discussion of the criteria for agent utility functions. The overall objectives of this paper are (1) to create a starting point for protocol engineering by future negotiation designers and (2) to enumerate the criteria and their measures that enable negotiation and allocation mechanisms to be compared objectively.

Automated planning is the field of Artificial Intelligence (AI) that focuses on identifying sequences of actions allowing to reach a goal state from a given initial state. The need of using such techniques in real-world applications has brought popular languages for expressing automated planning problems to provide direct support for continuous and discrete state variables, along with changes that can be either instantaneous or durative. PDDL+ (Planning Domain Definition Language +) models support the encoding of such representations, but the resulting planning problems are notoriously difficult for AI planners to cope with due to non-linear dependencies arising from the variables and infinite search spaces. This difficulty is exacerbated by the potentially huge fully ground representations used by modern planners in order to effectively explore the search space, which can make some problems impossible to tackle.

This paper investigates two grounding techniques for PDDL+ problems, both aimed at reducing the size of the full ground representation by reasoning over the lifted, more abstract problem structure. The first method extends the simple mechanism of invariant analysis to limit the groundings of operators upfront. The second method proposes to tackle the grounding process through a PDDL+ to classical planning abstraction; this allows us to leverage the amount of research done in the classical planning area. Our empirical analysis studies the effect of these novel approaches over both real-world hybrid applications and synthetic PDDL+ problems took from standard benchmarks of the planning community; our results reveal that not only the techniques improve the running time of previous grounding mechanisms but also let the planner extend the reach to problems that were not solvable before.

This is the second of two special issues focusing on the integration of artificial intelligence (AI) and operations research (OR) techniques for solving hard computational problems, with an emphasis on planning and scheduling. Both the AI and the OR community have developed sophisticated techniques to tackle such challenging problems. OR has relied heavily on mathematical programming formulations such as integer and linear programming, while AI has developed constraint-based search techniques and inference methods. Recently, we have seen a convergence of ideas, drawing on the individual strengths of these paradigms.

This paper attempts to assess the practical utility of nonmonotonic logic in diagnostic problem solving. We begin with a brief review of the main assumptions which motivate work in this area, and discuss two logic-based approaches which involve nonmonotonic arguments. Then we consider two recent proposals for the application of default logic to diagnosis, as well as a proposal based on counterfactual logic. In conclusion, we briefly compare these methods with other diagnostic reasoning paradigms found in the Artificial Intelligence literature.