The Frame of Reference Problem in the Design of Intelligent Machines

摘要

Criticisms of cognitive science and Al may often fail to be effective because they aren’t sufficiently grounded in computational modeling terminology and may even appear to be compatible with existing programs. For example, the current buzzword “situated” might just mean “conditional on the input data of particular situations”; hence all programs are situated. Moreover, the discourse of other intellectual traditions may appear incoherent to cognitive scientists; consider for example the claim that “representation must be based on interactive differentiation and implicit definition” (Bickhard & Richie, 1983). Experienced Al researchers believe that an engineering approach is essential for making progress on these issues. Perhaps the most important reason for recent progress and optimism about the future is the construction of alternative cognitive models as computer programs, the field’s agreed basis for expressing theories:• The Al-leaming community is focusing on how a given ontology of internal structures-the designer’s prior commitment to the objects, events, and pro­ cesses in the world-enables or limits a given space of behavior (e.g., the knowledge-level analyses of Dietterich, 1986; Alexander et al., 1986);• New robots (“situated automata”) demonstrate that interpreting a map of the world isn’t required for complex navigation; instead, maintaining a relation between an agent’s internal state and new sensations enables simple mechanisms to bring about what observers would call search, tracking, avoidance, etc. (Agre, 1988; Braitenberg, 1984; Brooks, chap. 8 in this volume; Rosenschein, 1985; Steels, 1989);• Neural networks, incorporating “hidden layers” and using back-propagation learning, provide a new means of encoding input/output training rela­ tionships, and are suggestive (to some researchers at least) of how sensory and motor learning may occur in the brain (Rumelhart, McClelland & the PDP Research Group, 1986).