The title of this book is a reflection of both the deep skepticism regarding the existence of intelligent machines and the exaggerations of their mental prowess. The skepticism usually comes from the philosophical community, and ironically, from many researchers in artificial intelligence (AI). The exaggerations find their origin predominantly in Hollywood, with their portrayals of killer robots and machines run amok. But they arise also, and again ironically, from many AI researchers. This usually occurs when new and promising approaches are attempted, and the initial excitement causes some investigators to impute qualities to their machines that are difficult for interested observers to accept sometimes. After a rather sustained period of enthusiasm, one usually finds a period of reflection on just what is being done. After the research is analyzed and understood, it is then usually designated as being a "program" or some other label equally as banal. Skepticism regarding its "intelligence" grows accordingly.
This roller coaster ride of confidence and skepticism is the result of the lack of an effective tool that would discriminate between intelligent and non-intelligent machines. The AI community is certainly aware of this, and this book can be viewed as part of the search tree for this discriminator. Based on only two of the articles that were studied by this reviewer, it is a work that should be perused by anyone interested in the history of artificial intelligence, but especially for those who are working diligently to find a workable definition of machine intelligence. It is fair to say that the book reflects the attitudes of those in the control theory community, which in fact might be its biggest virtue since that field stays close to what can actually be implemented in real machines in the field.
The article entitled "What Makes A Machine Intelligent" by C.W. de Silva, the editor of the book, holds that it is the ability to find approximate solutions to problems that is the characteristic sign of intelligence. Referring to approximation as a "soft" concept, intelligent machines the author says are part of the field of "soft computing", and after a short review of this field he presents a general framework for assessing the intelligence of a machine. Most importantly, he believes that intelligent machines are prevalent in the world today, and find application in the household, electronics, transportation, and manufacturing. Interestingly, the author puts quotation marks around the word `brain' when writing that there are many products that have a "brain." He is thus displaying reluctance (or caution) in claiming that there are machines that can "think". Why not say that machines have a "brain" if they indeed exhibit intelligence? Can a machine exhibit intelligence without a "brain"?
But it is the knowledge base of a machine that enables machine intelligence the author argues, in that manipulation of this knowledge base can result in effective decision-making and inferences. And also crucial to the author's view of intelligence is that it is non-reductionist: at the "atomic" level the intelligence of the machine cannot be detected, much as at the neuronal level one does not observe intelligence in humans. Therefore, he argues, it is the external behavior or characteristics that determines to a large degree whether the machine is to be designated as intelligent. He lists several characteristics that if exhibited by machines would be a sign of their intelligence. One of these includes pattern recognition, and via the use of connectionist networks, this has become a reality. Another is inference from incomplete information, and this has been realized using inductive logic programming. Still another is inference from qualitative or approximate information and this has been achieved with various approaches with qualitative reasoning "software. And machines can now exhibit a fair degree of commonsense, the Cyc project being a prime example of the integration of commonsense into machines. However, there are no machines as of yet that exhibit curiosity and that engage in creative invention, and the author is careful to note this in his article. The construction of machines that exhibit these characteristics is one of the major challenges of twenty-first century AI, but it is the opinion of this reviewer that it will be accomplished within the next two decades.
The general schematic that the author gives for an intelligent machine is interesting in regards to its domain specificity. The author argues that efficiency drives this specificity, and that such intelligent systems are to be favored over general purpose ones. If one is emphasizing applications, it is difficult to argue with this point. For example, if one were interested in a system that would intelligently manage a wide area network, then one would not want this system to engage in chess playing or musical composition, even if it were capable of doing so. Resource allocation and context thus governs to a large extent what kind of intelligent machine is to be deployed. General-purpose intelligent machines however have been the topic of much discussion of late, due to the insistence by many AI researchers that only such a capability should really be designated as intelligent.
One could designate the machine as discussed by the author as a triplet (M, C, K), where M is the `structural system', i.e. the physical hardware and devices, K is the knowledge base, and C is the "brain" that consists of various reasoning strategies acting on K. There is no reason a priori to insist that C and K are tied to an explicit location in the machine. They may indeed be distributed throughout the machine, and so systems such as the Internet could qualify as candidates for intelligent machines. He gives an explicit example of such a machine, namely one that is able to perform intelligent sensor fusion.
A very different and highly interesting outlook on machine intelligence is given in the article "Intelligent Machine to Modify or Adapt Human Behavior" by D.W. Repperger. The author's view on intelligence is unique in that he asserts that if the overall interaction between a human and machine is improved, over and above what can be accomplished when there is no interaction, one can say that the machine expresses intelligence. The machine has thus assisted the human in some way, and caused her to increase her knowledge base or abilities, or in fact induced an adaptation to her behavior. The author gives two examples illustrating his definition of machine intelligence, and interestingly, these examples illustrate the idea that a machine can be viewed as intelligent in one time frame and not in another. The machine must add "value" to the machine/human interaction, and a machine that was able to do this two decades ago may not be able to do it today. The utility of the human/machine interaction is therefore time-dependent and either may increase in time, if the machine evolves its knowledge base or the sophistication of its reasoning strategies, or decrease with time, this happening when the machine is unable to keep up with ever-increasing knowledge or whose reasoning strategies prove impotent when coping with contemporary problems.
The author then outlines a possible taxonomy for machine intelligence, requiring first that any measure of machine intelligence must have a corresponding analogy in humans or animals. This requirement is somewhat restrictive, since highly advanced machines might formulate new languages or methodologies to communicate among themselves, which may have no analog in human or animal communication. But it would be difficult to recognize when this did happen however, at least from the standpoint of humans, and so the author's requirement does carry practical weight. Another requirement is that when a human is given a certain task, it is better able to complete that task when interacting with the machine in question. Also, a machine may express intelligence for one task but not another. For example, a machine could manage a network quite superbly but fail miserably at chess. Fourthly, intelligence measures must always be relative, not absolute, i.e. whether a machine is intelligent depends on the context and historical epoch in which it is embedded. A machine could be unintelligent a few years ago but be intelligent today, due to its ability to evolve and meet new challenges, or conversely, due to disruptions in human societies that require them to use machines that they used to regard as practically useless. Lastly, the author believes that a quantitative measure of machine intelligence should be related to information theory, such as channel capacity. As an example of an intelligent machine, the author discusses a system that detects tracking error for jet pilots.
