Invited Lecture IV - August 14, 2012
AGI systems that are intended to model human-level intelligence need to cover a large variety of different reasoning abilities. Such systems should, for example, be able to draw deductive, inductive, and abductive inferences, they should cover uncertain and vague reasoning types and should be able to learn from (potentially noisy) data. Nevertheless, such classical reasoning and learning paradigms are not sufficient to describe the breadth of human-level intelligence. Quite often, natural agents perform inferences by analogical reasoning and similarity-based reasoning. Furthermore, they are able to solve problems creatively by blending two conceptual spaces. This lecture argues for the claim that the design of isolated and special formalisms for such reasoning tasks is fundamentally wrong for an AGI system. Instead of developing such isolated formalisms an integrated approach that incorporates all types of reasoning in one framework is proposed. More specifically, this lecture argues for the integration of certain cognitive mechanisms into AGI systems. This will be exemplified by the specification of Heuristic-Driven Theory Projection (HDTP) (Schwering et al., 2009), a framework designed for analogical reasoning and concept blending that is able to model also classical reasoning types. The lecture motivates the importance of cognitive mechanisms for general intelligence from a cognitive science perspective. The syntactic and semantic principles of HDTP will be introduced. Additionally, the application of this framework for solving complex cognitive tasks will be sketched.