on January 16, 2006
It's been about 50 years since the word Artificial Intelligence was coined. Since then there have been a number of television shows and movies about AI, but in real life AI has yet to produce a young boy to life an even quasi-normal life.
Behind the scenes however, research has been going on to develop the sub-systems needed as a foundation of AI. In this book the author describes what's going on in computers about such critical areas as vision, speech, taste, smell and so on.
The big problem, and what's covered in most of the book are what you might call the thinking components. How do computers think? How do they play games such as chess? Or one of the hot new items, play soccer. Then there are real problems like getting the computer to write fiction? Can a computer be programmed to transpose bits and bytes into thought, or love?
There have been a number of books lately on robotic activities you can do at home. This one is a description of the state of the art in the research labs around the world.
Throughout the last five decades, fed by both curiosity and military requirements, the design and construction of robots has occupied the time of many researchers, and involved the spending of hundreds of millions of dollars. In this book the author presents an overview of robotics for a semi-popular audience, beginning with a fairly detailed summary of the early history of artificial intelligence. It should be remembered that robotics is but one subfield of artificial intelligence, and that the latter field encompasses much more than the building of humanoid-looking machines. And interestingly, when one compares the research forty or even fifty years ago with what is going on at the present time, it is readily apparent that the differences are more of quality rather than quantity.
But intelligent machines do not have to take the form of humanoid robots. Hollywood and science fiction novels are partly responsible for this attitude, as are the philosophers, who insist upon the Turing test as being a genuine test for machine intelligence. It is evident when reading the book, especially the last part, that the author will not be convinced of the existence of intelligent machines until they do most, if not all, of the things that humans do. This includes the ability to make love, the ability to reproduce, the possession of legal rights, the possession of consciousness, and the ability to feel emotion and fall in love. A machine taking the form of a humanoid robot that was able to do all of things would certainly qualify as being intelligent. But there are many other types of machines, some of which exists today and are working in the field, that qualify as being intelligent, even though it is a different type of intelligence than what most humans are used to (or would acknowledge as such).
This observation raises another issue that is noticeably lacking in this book, as well as in the history of artificial intelligence in general. This issue involves the adoption of a quantitative definition of machine intelligence that will allow its measurement. If one is to judge the progress in artificial intelligence, it is necessary to define criteria, possibly informal, for assessing to what degree one machine is more intelligent or of higher quality than another. The criteria must also be able to distinguish an intelligent from a non-intelligent machine. The Turing test is not entirely suitable as a criterion, since it emphasizes, somewhat myopically and exclusively, human intelligence as being the most objective measure.
After careful study of the history of artificial intelligence, in this book and many others, as well as research papers, and through the development and practical use of `algorithms' that are deemed to be intelligent in some way, this reviewer arrived at an informal classification scheme for intelligent machines. Sometimes this scheme allows the quantitative measurement of machine intelligence, a `machine IQ' if you will, but usually it classifies machines according to what they can do, and to the degree that the machines require assistance from another machine (human or not).
For example, one could label a machine `Type-1' if it is an ordinary calculating machine, unable to learn or check its answers, or unaware of its environment. Type-1 machines are uninteresting from the standpoint of artificial intelligence research. A `Type-2' machine can find answers to domain-specific problems and check these answers according to standards given to it from another machine. Type-2 machines essentially need `tutors' or some kind of assistance to evaluate or continue learning. The chess playing machines described in this book, such as Deep Blue and Deep Thought, could be classified as Type-2 machines. The Pinkerton music-creating machine is also Type-2 as are the rule-based music-creating machines discussed in the book.
`Type-3' machines are able to check their answers to domain-specific problems and make judgments as to the quality of these answers, and do independently of any external standards. The Samuel checkers playing machine and the NeuroGammon and TD-Gammon backgammon playing machines described in this book could be classified as Type-3 machines, as would the `metagame' machines that can learn how to play a game given only the rules. Also Type-3 is the bridge-playing COBRA machine, and the Poki poker-playing machine, the Thaler Creativity Machine, the BRUTUS storytelling machine, all of which are discussed in the book.
A `Type-4' machine is one that is able to judge the quality of its answers to domain-specific problems and then propose theories or explanations that subsume these problems. Type-4 machines are thus machines that one could use to conduct scientific research for example. The EMI music-making machine discussed in the book is a Type-4 machine, due to its ability to analyze the structure of the music presented to it, and then extract the composer's style from it. Type-4 machines have been used in automated drug discovery, although this use is not discussed in this book.
Next are the `Type-5' machines, which are able to solve problems in more than one domain, but with their interest in solving these problems is instigated by an external inquirer, i.e. they do not possess any innate curiosity. The `commonsense reasoning' machines of Cycorp, Inc, which are discussed in the book, are examples of Type-5 machines. It is their ability to solve problems in more than one domain that makes Type-5 machines of great interest to many in the artificial intelligence community. Many in fact do not believe a machine is truly intelligent unless it can think in more than one domain.
A `Type-6' machine can express curiosity and creativity, can solve problems without any external instigation, and can develop theories or explanations around these problems. The author discusses several types of machines in the book that could be classified as Type-6, if one omitted the ability to find solutions without being instigated by an external machine or human.
Lastly, there are `Type-7' machines, which can self-manage and self-replicate, and are also Type-6. Self-replication is discussed in the book, but there are no machines to date that are Type-7.