Smarts are in the architecture

摘要

A list of desiderata for an autonomous agent architecture includes decentralized, yet coherent coordination of behaviors, easy interface for a designer and end user, distinction of knowledge and skills, easily scalable, quantifiable measures of success in interaction in the real world and quantifiable learning metrics, ability to learn new knowledge and skills, and reliability. We present our architecture which aims to meet this list. Several instantiations including a Lego robot instance of our architecture is discussed in some detail. We will explore general questions in hope of working towards quantifiable parameters that guide architectural choices. What is an architecture? An architecture is not a list of choices for hardware and software. An architecture is also not a list of programming paradigms, or hardware philosophies. As a superset, an architecture includes the principles of how an agent perceives its environment, how an agent has tasks and goals, how an agent has attitudes about objects and agents, how an agent has internal motivations, and how it attempts goal satisfaction. An architecture cannot be understood by an agent observer who only examines inputs and outputs (or the function of an algorithm). So in that regard, an architecture is less of an engineering endeavor since an engineered system can be purely judged based on its inputs and outputs (or execution of a computer program). Judging an architecture is more like judging a programming style. Just as a computer program style is best judged by an expert programmer, an architecture is best judged by a person who builds and runs autonomous agents.