一个维特根斯坦主义者眼中的框架问题  

The Frame Problem from a Wittgensteinian Point of View

作　　者：徐英瑾[1] 

机构地区：[1]复旦大学哲学学院

出　　处：《逻辑学研究》2011年第2期93-136,共44页Studies in Logic

基　　金：2006年国家社科基金项目"维特根斯坦哲学视野中的人工智能问题"(06CZX011)的资助

摘　　要：框架问题的核心命意乃是:在对智能体的行动进行表征之时,如何能够在有效地表征出行动的效果之时,避免对于其非效果的表征?对于符号主义路数的人工智能研究来说,框架问题的确非常棘手,因为它的存在使得对于行动的知识刻画必须引入大量的框架公理,并由此使得系统的计算负荷剧增。从历史上看,人工智能界赖以解决这个问题的基本路径主要有两条。第一条是诉诸于"关于惯性的常识律"或缺省逻辑,即在行动表征中预设:没有被表征式牵涉到的"世界其余部分"都将保持惯性。另一条路径则采纳了海耶斯的"素朴物理学"表征方案,即试图把日常物理学知识全部用谓词逻辑公理化,让系统自动推算出哪些动作会导致哪些结果。本文将论证,这两条道路其实都走不通。在此基础上,本文将从维特根斯坦的《论确定性》中寻找灵感,建立一个关于信念网动力学模型——在这个模型中,某些知识节点的激发只会导致某些与之毗邻的节点的激发,因此,对于特定问题求解路径的表征,就能够避免对于一个更大范围内的知识网络的牵涉。框架问题也有望借助于这个模型得到解决。最后,我将利用"非公理推演系统"(NARS)为这个模型提供一种尽量逼真的工程学模拟。The basic worry condensed in the so-called＂Frame Problem＂in Artificial Intelligence （AI）is：When an artificial agent＇s actions are being represented in forms of programs, how can a programmer of this system represent the desired effects of these actions without representing all of their none-effects？（He should be able to do this,otherwise he would introduce too many frame axioms into the action-representation of the very agent, but this will definitely render the computing budget of the whole system too big in the real time.）There are two basic approaches to this problem in the literature：First,by appealing to the＂common sense law of inertia＂or default logic,which can be explicated by the metaphysical assumption that the untouched part of the world is always inertial, a programmer is supposed to be able to avoid representing those undesired non-effects; second,by appealing to Hayes＇＂Naive Physics＂project,which aims at a systematic formalization of human＇s common sense of the physical world in terms of predicate logic,a programmer is also supposed to be able to avoid representing those frame-axioms,given that the built-in knowledge stored in the computers has already constrained the logical space of representing the action-triggered effects.However,I will try to argue that neither approach is promising.Then,I will sketch a Wittgenstein-inspired cognitive model on belief-revision,i.e.,an experience-revisable semantic network in which the activation of some nodes,representing causes,will only lead to that of other nodes,representing the desired effects.Finally,I will provide an engineering approximation to this model with the help of Non-Axiomatic Reasoning System（NARS）.

关 键 词：框架问题 关于惯性的常识律 素朴物理学 缺省逻辑 维特根斯坦 非公理推演系统 

分 类 号：B81[哲学宗教—逻辑学]