DFG-Heisenbergproject: Formalization, modeling and implementation of a unified neuro-cognitive theory of reasoning

Reasoning is at the core of many cognitive systems and relevant for artificial intelligence and cognitive science. To develop a cognitive reasoning framework (e.g., for a cognitive agent) we need to have a cognitive theory and model that adequately reflects human reasoning. However, today a great variety of competing psychological reasoning theories exists. Recently a meta-analysis of syllogistic reasoning theories revealed that all of them significantly deviate from human data.To develop a comprehensive cognitive theory, a formal and computational assessment of existing theories is necessary. This is challenging since most theories are only informally described, they are restricted to a specific domain and level of analysis (i.e., behavioral, algorithmic, or neural level), and no general benchmark set exists.A comprehensive analysis of reasoning theories requires a multi-methodological approach combining techniques from artificial intelligence like cognitive modeling and knowledge representation and reasoning with empirical investigations. Recently developed multinomial processing tree models enable a quantitative comparison of different reasoning theories. Cognitive architectures allow for an additional integration of working memory and connect symbolic theories to fMRI-findings.The objective of this project is such a computational analysis of reasoning theories and to define and implement a comprehensive and domain-independent neuro-cognitive theory of human deductive reasoning. This requires (i) a thorough formal and algorithmic analysis and implementationt o make cognitive theories usable and to analyze their predictive power, (ii) an assessment of theories considering a uniform benchmark set and different levels of analysis (behavioral, algorithmic, and neural); and (iii) extracting common factors about mental representations and operations towards a comprehensive theory of neuro-cognitive reasoning. Such a formal and algorithmic approach makes the theory readily available for building cognitive agents that can interact naturally with humans in reasoning processes.