Understanding Control at the Knowledge Level

What is it that unifies the control task in all its manifestations, from the thermostat to the operator of a nuclear power plant? At the same time, how do we explain the variety of the solutions that we see for the task? I propose a Knowledge Level analysis of the task which leads to a task-structure for control. Differences in availability of knowledge, the degree of compilation in the knowledge to map from observations to actions, and properties required of the solutions together determine the differences in the solution architectures. I end by discussing a number of heuristics that arise out of the Knowledge Level analysis that can help in the design of systems to control the physical world. What Is the Knowledge Level? By now most of us in AI know about the Knowledge Level proposal of Newell [Newell, 81]. It is a way of explaining and predicting the behavior of a decision-making agent without committing oneself to a description of the mechanisms of implementation. The idea is to attribute to the agent a goal or set of goals and knowledge which together would explain its behavior, assuming that the agent is abstractly a rational agent, i.e., one that would apply an item of relevant knowledge to the achievement of a goal. Imagine the following conversation between a guest and a host at a house party: G: Why does your cat keep going into the kitchen again and again? H: Oh, it thinks that the food bowl is still in the kitchen. It doesn’t know I just moved it to the porch. The host attributes to the cat the goal of satisfying its hunger, and explains its behavior by positing a (mistaken) piece of knowledge that the food is in the kitchen. That the cat would naturally go to the kitchen under these conditions seems reasonable to the host and presumably to the guest. When people talk this way, they are not asserting that the neural stuff of the cat is some kind of a logical inference machine working on Predicate Calculus expressions. It is simply a useful way of setting up a model of agents and using the model for explaining their behavior. The attributions of the goal and knowledge can be changed on the basis of further empirical evidence, but the revised model would still be in the same language of goals and knowledge. The Knowledge Level still needs a representation, but this is a representation that is not posited in the agent, but one in which outsiders talk about the agent. Newell thought that logic was an appropriate representation for this purpose, leaving open the possibility of other languages also being appropriate in some circumstances. Newell used the phrase "Symbol Level," to refer to the representational languages actually used to implement artificial decision agents (or explain the implementation of natural agents). Logic-based languages, Lisp and FORTRAN, neural net descriptions, and even Brook’s subsumption architectures are all possible Symbol-Level implementations for a given Knowledge Level specification of an agent. The Control Problem at the Knowledge Level Consider the following devices and control agents: the thermostat, the speed regulator for an engine, an animal controlling its body during some motion, the operator of a nuclear power plant, the president and his economic advisors during the task of policy formulation to control inflation and unemployment, and a major corporation planning to control its rapid loss of market share. All these systems are engaged in a 19 From: AAAI Technical Report FS-94-03. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved.