Computational Rationality

在智能行为的计算研究中，有几种现象是孤立的，其中最重要的是学习、推理和计划。已经发现普遍存在的障碍是，各种现象的形式化常常彼此不一致，甚至更常见的是，它们在计算上难以处理。这个研究项目的主要目的是开发一个计算框架和一些算法，这些框架和算法将构成这个问题领域的统一方法。本文提出的出发点是，可能近似正确（PAC）学习的概念为系统的行为性能提供了一个定量的标准。这些用于计算学习理论的分析方法被扩展到更广泛的现象类别，包括各种推理和规划的概念。

In computational studies of intelligent behavior several phenomena have been isolated, among the most important of which are learning, reasoning, and planning. Obstacles that have been found pervasive are that the formalization of the various phenomena are often inconsistent with each other, and, even more often, that they are computationally intractable. This main thrust of this research project is to develop a computational framework and some algorithms that would comprise a unified approach to this problem area. The proposed starting point is the idea that the concept of probably approximately correct (PAC) learning provides a quantitative criterion on the behavioral performance of a system. These analytic methods used in computational learning theory, are extended to a wider class of phenomena, including various notions of reasoning and planning.