The Eighth International Conference on Principles of Knowledge Representation and Reasoning (KR2002) will be held in Toulouse, France from 22 to 25 April 2002. KR2002 will be held in conjunction with AIPS'02.
Explicit representations of knowledge manipulated by inference algorithms provide an important foundation for much work in Artificial Intelligence, including natural language dialogue systems, high level vision, robotics and other knowledge based systems. The KR conferences have established themselves as the leading forum for timely, in-depth presentation of progress in the theory and principles underlying the representation and computational manipulation of knowledge. The traditional very high standard of papers will be maintained at KR2002.
KR2002 will have the following associated workshops:
KR2002 will be held at the Pierre Baudis Convention Center, in downtown Toulouse. More information about Toulouse can be obtained from its tourist office.
Traditional symbol based theories of induction and concept formation have built on the following two assumptions (among others): (1) There is a set of primitive properties (predicates), from which new concepts are constructed and correlations between properties are determined. (2) Correlations between predicates are described by conditional probabilities. However, the first assumption does not answer the question of how we can learn the primitive predicates in the first place. To handle this problem, machine learning algorithms or models based on artificial neuron networks have been used instead. The second assumption has been motivated within the Bayesian tradition as being the most rational way of determining correlations between properties. The use of conditional probabilities is motivated by the principle of maximal entropy (Williams 1980). However, by refering to maximal entropy all instances of a concept are treated as unrelated. On the other hand, in human concept formation and inductive reasoning, judgments of similarity are essential. Similarity cannot be handled in a natural way in terms of Bayesian conditional probablities. I will present a model, developed together with Christian Balkenius, of how (higher order) similarities can be used in induction and concept formation. The model is based on a layered structure of self-organising maps (Kohonen 1995). The model will be related to the conceptual spaces developed in my recent book (Gärdenfors 2000). It will be shown how the model can handle some facts concerning human inductive reasoning that have been problematic for models based on conditional probabilities. I will also connect the model to the theory of case based reasoning proposed by Gilboa and Schmeidler.
For many years the knowledge representation community has worked with a set of assumptions that have led to a solid body of work - formal, clean, and largely irrelevant to the world at large. Expressivity, consistency and decidability have been the guiding principals of the field, with little worry about actual performance, scaling or usability issues. In short, the field has flourished on theoretical elegance and incremental results, with little perturbation over the past decade or so.
Recently however, a new wind has been blowing through the knowledge representation world. The World Wide Web has been seeing a growing demand for semantics! Large collections of web pages, images and collections need organizing principals for management. Databases and web services must be able to advertise capabilities and find each other. XML datasets and schemas need to be linked in scalable ways that allow a WEB of semantic information to emerge.
Unfortunately, most of those demanding knowledge representation on the web have different goals than those of the KR community. Scalability and performance trump decidability, ease of use fights with expressivity in language design, and consistency is impossible to guarantee, if it is even desirable. In short, the world is demanding KR, but it refuses to agree to the methodology of our community!
Other researchers, such as those in hypertext and information retrieval have seen their fields changed beyond recognition as the web reached into their communities. KR is next! but are we up to the challenge?? In this talk, I will describe the needs of this new world of web semantics, and challenge the KR community to rise to the historic opportunity to help address them.
The main task of a soccer player is to score goals. However, there are moments in life when questions like the following become relevant: Is the ball I am seeing a hallucination or is it real? Should I revise my beliefs about where the ball is? And if so, what is the next action I should execute? Would this action be to the benefit of my team? In the talk I will address these questions and show how one can create a successful robotic soccer team by giving the right answers.