構成的情報学とAI(AIフロンティア論文)

抄録

As computer scientists, we have been trained in the methodology of natural science, which is analytic in its essence. Informatics, and particularly Artificial Intelligence, is not an analytic discipline. It is required to establish a constructive methodology. The gap between two disciplines is larger than as it is generally conceived. It is important to recognize the difference and establish a new methodology. One of the motivating findings to write this paper is that western and eastern views of the world are quite different. The former assumes existence of an observer that is independent from the observed system and the view point is unconsciously fixed to the observer rather than the system. This paper argues that eastern view of the world is essential in search for the new methodology. In designing a system, that requires intelligence in particular, it is essential to take the agent's view that is inside the system rather than God's eyes' view separated from the system. The direct consequences of this change allow us to discuss about and gives design principles on the following issues: (1) interaction between an AI system and the AI researcher, (2) partiality of information, (3) situatedness, (4) multi-layered complex systems, (5) emergence and constructive methodology, and (6) a theory as a good story. We need a world view that allows locally consistent but globally inconsistent systems. There are researchers to claim that intelligence is built on top of ad-hoc miscellany of many special purpose modules rather than on top of small coherent collection of basic algorithms. In this paper, I propose a new research methodology for constructive sciences. It includes analytic methodology as its part. However, the methodology is by no means complete. It is a first step toward the completion of the constructive methodology, which may require joint effort of the community for more than a decade.