In this installment, I have invited the author of a new text on AI to present their motivations and perspectives on teaching AI

Deepak Kumar

Understanding Intelligence
by Rolf Pfeifer and Christian Scheier
MIT Press (1999)

Introduction
The target audience of Understanding Intelligence is students of artificial intelligence who are not only interested in clever algorithms but in understanding natural forms of intelligence and principles of intelligent behavior in general. The book discusses the synthetic methodology that can be characterized as "understanding by building". It consists of three components: (1) modeling certain aspects of biological systems, (2) abstracting general principles of intelligent behavior, and (3) applying these principles to the design of intelligent systems.

Modeling Aspects of Biological Systems
Many examples of biological systems that researchers have tried to model can be found in the literature: foraging behavior; path integration, visual navigation in ants and honey bees; schooling and predator avoidance behavior in fish; implicit learning in rats exploring a maze; the emergence of hierarchies in groups of chimpanzees involved in local dominance interactions; human infants learning to make distinctions in the real world; adults recognizing objects under various viewing angles, distances, and lighting conditions; experts performing medical diagnosis; mathematicians proving a theorem; and, of course, grand masters playing chess. Whereas biologists or psychologists might be satisfied with such a model, artificial intelligence researchers or cognitive scientists will want to know about the general principles.

Abstracting General Principles of Intelligent Behavior
In the classical view of artificial intelligence, the general principles dealt mostly with symbol processing and computational architecture. In more recent approaches, where embodiment plays an important role, the principles that have been suggested are more strongly related to the interaction with the real world as it is mediated by the body of the agent. Here are, very briefly, a few examples: One principle asserts that we must not look at the agent in isolation but must define its ecological niche, its tasks, and at the types of interactions of the agent with its environment.

Another principle, inexpensive design, states that these interactions can be exploited in the design of an agent. A beautiful illustration of this principle is Ian Horsewill's robot Polly. In the early 1990s Polly gave tours of the MIT AI Lab. It's camera was slightly tilted downwards so that more distant objects were higher up on the y-axis in the image - an inexpensive way of visually detecting the nearest obstacles.

The principle of sensory-motor coordination was inspired by John Dewey, who, as early as 1896, had pointed out the importance of sensory-motor coordination for perception. This principle implies that through sensory-motor coordination, through coordinated interaction with the environment, an agent can structure its own sensory input. In this way, correlated sensory stimulation can be generated in different sensory channels --an important prerequisite for perceptual learning and concept development.

Another principle has its origins in the work of Rodney Brooks, who introduced into AI research the idea of embodiment and the subsumption architecture. According to the principle of parallel, loosely coupled processes, intelligence emerges from a large number of parallel processes that are only loosely coupled and are mostly coordinated through the interaction with the environment. An example is an insect walking: coordination of the individual legs is achieved not only through neural connections but also the environment. Because of the body's stiffness and its weight, if one leg is lifted, the force on all the legs changes instantaneously, a fact that is exploited by the leg coordination system in the insect.

Another principle that can be abstracted from various models is the complexity of the relationships among the sensory, motor and neural control systems. In intelligent systems --especially naturally intelligent systems and artificial systems --given a particular task, there is always an "ecological balance" in the complexity of the sensory system, the motor system, and the neural substrate. Additional principles deal with redundancy, value, learning, and self-organization (which are not further discussed here). Note that in the classical perspective most of these principles do not apply because it is limited to the computational world.

Applying the Principles to the Design of Intelligent Systems
Once the principles have been established, they can be applied to designing intelligent systems and practical applications. The book has many case studies that illustrate the various principles.

Content of a Course on New AI
We have been teaching a cognitive science oriented class, called"An Introduction to New AI" (also called "Embodied AI") for a strongly cross-disciplinary audience, including computer scientists, engineers, psychologists, biologists, neuroscientists, mathematicians, and physicists. Understanding Intelligence, the textbook we use, outlines the issues involved in the study of intelligence and the classical paradigm and its major problems. It then sketches a framework for embodied artificial intelligence and fundamental topics to consider when designing and analyzing intelligent systems. An example is the frame-of-reference problem which was discussed by a number of authors, most notably Herbert Simon using his famous "ant on the beach." From an observer's perspective, the ant's path is highly complex, but the mechanisms underlying its behavior might in fact be very simple. For example, behavior rules such as "if obstacle on left turn right; if obstacle on right, turn left", operate in the neural substrate of the ant. The behavior of the ant cannot be reduced to its internal neural mechanism because behavior is always an interaction with the real world. Increasing the size of the ant by a factor of 1000, but using the identical neural program, the ant would, in exactly the same environment, describe a path that would be much straighter. Other general principles such as the varous time scales involved in explaining behavior and designing intelligent systems, are also discussed.

The book gives an overview of the various approaches to designing and explaining intelligent systems, including adaptive neural networks (i.e. networks that do not need to be trained with defined training and test sets, and that can learn as they are performing in the real world), Braitenberg vehicles, behavior-based robotics and the subsumption architecture, and artificial evolution. In addition, it briefly summarizes the behavioral economics, dynamical systems, and schema-based approaches to designing intelligent systems. The next five chapters summarize a set of abstract principles, called "design principles of autonomous agents", some of which were illustrated earlier. One chapter in particular should appeal to psychologists and cognitive scientists, who intend to apply principles from embodied AI (or embodied cognitive science) to a high-level cognitive phenomenon, namely memory. Although memory is considered a phenomenon of "high-level cognition", there are many direct links to embodiment. Another part of the book discusses general design issues and evaluation of models. The last part is dedicated to a re-assessment of artificial intelligence from an embodied perspective. Although, the field of "New Artificial Intelligence" has been around only for about 15 years, many exciting research issues and teaching ideas have emerged from it.

Web Resources
If you are interested in teaching a class, the book has a companion web site http://www.ifi.unizh.ch/~pfeifer/mitbook/ that is being continually developed. The website gives a number of programming examples. A simulator on the site provides some insight on various problems. We need an "agent simulator", one that can simulate the actual interactions of an agent with its environment. The website also contains comments on robot implementations. Most experiments discussed in the book can be performed with a generic robot architecture: a robot with two motors and a number of IR sensors, and perhaps a few light sensors.

For those interested in the theoretical background and research issues, you can use chapters 11 through 17 as a basis for discussions: they discuss the design principles, present evidence from various disciplines, demonstrate detailed case studies, and outline research topics.

Students from different disciplines who have attended this class have been excited; we plan to teach it again in the upcoming winter term.