Dimensions of Interaction

Abstract

gent, but deaf, blind, and paraplegic agents. The assumption was that a human agent would formulate the problem in a previously defined language understandable to the agent. The problem statement would include background knowledge, a description of the state of some world, operators to use in that world, and a description of a desired state (a goal). The agent would use (at the knowledge level) logical reasoning to determine a sequence of operations that could be performed to achieve the desired goal, and a more complete description of the resulting desired state. Although this style of building AI systems is still both useful and prevalent, there has been increasing recognition of the importance of designing and analyzing systems that do not make these isolation assumptions. In this article, I want to focus on systems consisting of active agents with multiple goals, communicating among themselves, and interacting Ten years ago, at the first AAAI conference, Alan Newell (1982), in his presidential address, focused on understanding the then dominant paradigm for artificial intelligence: the writing of symbolic reasoning programs for an agent that would act rationally to achieve a goal. He distinguished the “knowledge level” from the “symbol level” description of a system. The former encompasses the knowledge and conditions necessary for an agent to solve a problem. It is only when the knowledge level is reduced to a symbol level description that issues of implementation get considered. Newell’s analysis postulated a single agent, with fixed knowledge and a specified goal. The agent was disconnected from the world, with neither sensors nor effectors, and more importantly with no connection to other intelligent goal -dr iven a g e n t s . R e s e a r c h results in AI consisted pr imar i ly in the determination of AAAI-90 Presidential Address