Abstract
This collection of eleven PhD dissertation abstracts, originally published in The Knowledge Engineering Review (2008), provides a comprehensive overview of the evolving landscape of Artificial Intelligence and Knowledge Engineering. The research spans diverse domains—from autonomous planning under uncertainty to spatial data mining and human-computer interaction. A central theme across these works is the transition from rigid, theoretical frameworks to flexible systems capable of handling the "messiness" of real-world environments, including incomplete information, massive datasets, and human behavioral nuances.
Introduction
The mid-2000s marked a pivotal era for AI research, characterized by a shift toward practical application in uncertain environments. This compilation serves as a "time capsule," highlighting how doctoral researchers addressed the limitations of early AI systems when faced with the complexity of the physical and digital world. Key challenges included planning with imperfect knowledge, filtering redundant patterns in big data, and modeling human reasoning for more natural interaction.
Planning Under Uncertainty
A significant contribution to the field of automated planning is presented by Tuan Chi Le, who developed a state-based regression function for domains where agents lack complete information. By employing a "0-approximation" semantics, Le utilized three-valued logic (true, false, and unknown) to represent states. This approach, while potentially incomplete compared to full semantics, significantly reduces computational complexity. Le’s work demonstrates that a contingent planner using this regression function remains sound and complete relative to progression-based solutions, bridging the gap between theoretical rigor and practical efficiency.
Knowledge Discovery and Spatial Mining
As datasets grew in size and complexity, the challenge shifted from finding any pattern to finding meaningful ones. Vania Bogorny addressed this in the context of geographic databases. Traditional spatial association rule mining often produces "common sense" results (e.g., "cities are near roads"). Bogorny introduced early pruning strategies to remove these well-known dependencies, allowing for the discovery of truly novel insights. Similarly, in the medical domain, Vladutu utilized Self-Organizing Maps (SOMs) and clustering to analyze biomedical signals, such as myocardial ischaemia records, to identify critical diagnostic patterns.
Human-AI Interaction and Perception
The collection highlights a move toward more intuitive human-computer interfaces. Tevfik Metin Sezgin explored sketch recognition, arguing that the temporal context—how a user draws a symbol—is as vital as the final image. By using Multiscale Stochastic Models to capture "digital ink" signatures, Sezgin achieved significant improvements in recognition rates for electronic circuit diagrams. In the realm of education, Suebnukarn developed COMET, a medical tutoring system that uses Bayesian networks to model student reasoning, providing a framework for automated, intelligent pedagogical support.
Multi-Agent Systems and Robotics
The management of complex systems also saw innovation. Ronnie Johansson proposed a shift from "sensor management" to "perception management." By applying economic bargaining theory, Johansson created a framework where multi-agent systems "negotiate" for information, optimizing the collective perspective of the agents rather than just the hardware output. In robotics, Neto addressed the problem of novelty detection using "Grow-When-Required" (GWR) networks, allowing autonomous inspection robots to identify and react to unexpected changes in their environment in real-time.
Theoretical and Evolutionary Frameworks
Foundational logic and evolutionary biology also informed the research. Elizabeth Black introduced a generative framework for argumentation-based enquiry, enabling agents to reach conclusions through structured, dialectical dialogue. On the engineering side, Kyung-Joong Kim applied "speciation algorithms" to automatically modularize complex systems, mimicking biological evolution to overcome design bottlenecks and improve system scalability.
Conclusion
The dissertations summarized in this collection reflect a maturing field. By embracing uncertainty through three-valued logic, pruning redundant data in spatial mining, and incorporating temporal signatures in sketch recognition, these researchers moved AI closer to real-world utility. While many of the systems were domain-specific, they established the formal proofs of soundness and the algorithmic foundations necessary for the more robust, general-purpose AI systems of the following decade.
Summary updated