Engineering Intelligent Systems

The book Engineering of Mind: An Introduction to the Science of Intelligent Systems (Albus and Meystel 2001) outlines the main streams of research that we believe will eventually converge in a scientific theory that can support and bring about the engineering of mind. We believe that our research approach can enable the design of intelligent systems that pursue goals, imagine the future, make plans, and react to what they see, feel, hear, smell, and taste. We argue that highly intelligent behavior can be achieved by decomposing goals and plans through many hierarchical levels, with knowledge represented in a world model at the appropriate range and resolution at each level. We describe how a high degree of intelligence can be achieved using a rich dynamic world model that includes both a priori knowledge and information provided by sensors and a sensory processing system. We suggest how intelligent decision-making can be facilitated by a value judgment system that evaluates what is good and bad, important and trivial, and one that estimates cost, benefit, and risk of potential future actions. This will enable the development of systems that behave as if they are sentient, knowing, caring, creative individuals motivated by hope, fear, pain, pleasure, love, hate, curiosity, and a sense of priorities.

We believe that this line of research on highly intelligent systems will yield important insights into elements of mind such as attention, gestalt grouping, filtering, classification, imagination, thinking, communication, intention, motivation, and subjective experience. As the systems we build grow increasingly intelligent, we will begin to see the outlines of what can only be called mind. We hypothesize that mind is a phenomenon that will emerge when intelligent systems achieve a certain level of sophistication in sensing, perception, cognition, reasoning, planning, and control of behavior.

There are good reasons to believe that the computing power to achieve human levels of intelligence will be achieved within a few decades. Since computers were invented about a half-century ago, the rate of progress in computer technology has been astounding. Since the early 1950s, computing power has doubled about every three years. This is a compound growth rate of a factor of ten per decade, a factor of 100 every two decades. This growth rate shows no sign of slowing, and in fact, is accelerating: during the 1990s, computing power doubled every 18 months — a factor of ten every five years. Today, a typical personal computer costing less than $1000 has more computing power than a top-of-the-line supercomputer of only two decades ago. One giga-op (one billion operations per second) single-board computers are now on the market. There appears to be no theoretical limit that will slow the rate of growth in computing power for at least the next few decades. This means that within ten years, a relatively inexpensive network of ten single-board computers could have computational power approaching one tera-ops (one trillion, or 1012 operations per second). Within twenty years, ten single-board computers will be capable of 1014 operations per second. This is equivalent to the estimated computational power of the human brain (Moravec 1999). Thus, it seems quite likely that within two decades, the computing power will exist to build machines that are functionally equivalent to the human brain.

Of course, more than raw computing power is necessary to build machines that achieve human levels of performance. But the knowledge of how to utilize this computing power to generate highly intelligent behavior is developing faster than most people appreciate. Progress is rapid in many different fields. Recent results from a number of disciplines have established the foundations for a theoretical framework that might best be called a "computational theory of mind." In our book, Meystel and I have organized these results into a reference model architecture that we believe can be used to organize massive amounts of computational power into intelligent systems with human-level capabilities. This reference model architecture consists of a hierarchy of massively parallel computational modules and data structures interconnected by information pathways that enable analysis of the past, estimation of the present, and prediction of the future.

This architecture specifies a rich dynamic internal model of the world that can represent entities, events, relationships, images, and maps in support of higher levels of intelligent behavior. This model enables goals, motives, and priorities to be decomposed into behavioral trajectories that achieve or maintain goal states. Our reference architecture accommodates concepts from artificial intelligence, control theory, image understanding, signal processing, and decision theory. We demonstrate how algorithms, procedures, and data embedded within this architecture can enable the analysis of situations, the formulation of plans, the choice of behaviors, and the computation of current and expected rewards, punishments, costs, benefits, risks, priorities, and motives.

Our reference model architecture suggests an engineering methodology for the design and construction of intelligent machine systems. This architecture consists of layers of interconnected computational nodes, each containing elements of sensory processing, world modeling, value judgment, and behavior generation. At lower levels, these elements generate goal-seeking reactive behavior; at higher levels, they enable perception, cognition, reasoning, imagination, and planning. Within each level, the product of range and resolution in time and space is limited: at low levels, range is short and resolution is high, whereas at high levels, range is long and resolution is low. This enables high precision and quick response to be achieved at low levels over short intervals of time and space, while long-range plans and abstract concepts can be formulated at high levels over broad regions of time and space.

Our reference model architecture is expressed in terms of the Realtime Control System (RCS) that has been developed at the National Institute of Standards and Technology and elsewhere over the last 25 years. RCS provides a design methodology, software development tools, and a library of software that is free and available via the Internet. Application experience with RCS provides examples of how this reference model can be applied to problems of practical importance. As a result of this experience, we believe that the engineering of mind is a feasible scientific goal that could be achieved within the next quarter century.

0 0

Post a comment