13 of 14 people found the following review helpful:
5.0 out of 5 stars
A clear overview of the subject, July 30, 2006
This review is from: Commonsense Reasoning (Hardcover)
Occupying the time of many researchers over four decades, the goal of developing a machine that can reason in many different domains is finally reaching fruition. Progress in machine intelligence up until the last few years has been limited to abilities in a single domain, such as chess, checkers, backgammon, medical diagnostics, network management, and so on. Many researchers have held that the only way to eliminate this domain-specificity is to enable commonsense reasoning in machines. Humans of course can think in many domains. A chess master for example can also be a good writer or a good network engineer, and this is the case since any common elements in chess, writing, or network engineering can be recognized and exploited by humans. The integration and automation of commonsense reasoning in machines is challenging, but in this book an approach is outlined that has shown promise. The book of course is targeted to specialists and students in artificial intelligence, but it could also be accessible to cognitive psychologists, linguists, mathematicians, and others who are interested in the subject matter. The book is not only a theoretical discussion, for it discusses and points to practical tools that can be downloaded and used by the reader to illustrate the main issues in the book.
As the author defines it in the first paragraph of the book, commonsense reasoning as a process that takes information about a particular scenario and then makes inferences about other aspects of this scenario based on general knowledge of how the world operates. He discusses the main issues that arise in commonsense reasoning, such as the need for representing commonsense knowledge, the ability to reason about events, some of which may be concurrent or nondeterministic, and how to deal with space. Of particular importance in this discussion is the `commonsense law of inertia', which essentially makes the "intuitive" point that things will say the same unless they are affected by an event. However, this law is "violated" in many cases, the author discussing a few of these, and so a system for commonsense reasoning must be able to `release' certain `fluents' from this law.
Central to the book is the construction of the `event calculus', which is a system for representing commonsense knowledge and for implementing three types of reasoning abilities, namely `temporal projection', `abduction', and `postdiction.' The fundamental notion of course is the `event', which is an action taking place in the world. Also fundamental is the notion of a `fluent' that represents a property in the world that varies with time. All knowledge is represented declaratively in the event calculus, in order that it can be implemented in the logic programming paradigm, and it is based on a `many-sorted' extension of first-order predicate logic. Also central to the event calculus is the `timepoint' which represents an instant in time. The `discrete event calculus' is a version of the event calculus where the timepoint is restricted to the integers. Both of these systems are axiomatized early on in the book. Of particular interest and very important to commonsense reasoning is the notion of nonmonotonic reasoning and the accompanying notion of circumscription. That commonsense reasoning must be at times nonmonotonic is obvious, since certain premises held at one time may have to be altered when new information presents itself at a later time.
The event calculus is a particular formalization of commonsense reasoning, and for this to work over more than one domain it is necessary to describe formally the notion of a domain. The author calls this a `domain description' and he endeavors to make it as flexible or "elaboration tolerant" throughout the book in order that the event calculus is able to handle novel situations as they arise.
The extensive discussions and characterization of the event calculus throughout the book, as well as the examples and extensive references that are given, give the reader a solid understanding of it. Many of the concepts in the event calculus are similar to ones that are found in the other systems of artificial thought. One of these is the `triggering' of events, which is used in discrete event simulation for example, and where an event occurs if a particular condition becomes true. Events whose occurrence depends on something happening are of course very common in the real world, and any effective implementation of commonsense reasoning will have to deal with them. The challenge is to design a system that does not require all the pre-conditions be put in by hand. The axioms for triggering, along with the reasoning patterns that are used by the machine (prediction, abduction, and postdiction), should be sufficient for dealing with events that are contingent on other actions taking place. If they had to be anticipated individually, the number of statements to this effect would proliferate beyond measure, making the commonsense reasoning system extremely brittle and useless. In addition, and the author discusses this in some detail, repeated triggering must be suppressed. This is accomplished by incorporating additional trigger and effect axioms in the system.
The most interesting issue discussed in the book concerns the implementation of the mental states of `agents' in the system of commonsense reasoning. It is in this discussion that the author first introduces the concept of an agent, which he defines as an entity that engages in purposeful actions in the world. An agent can have beliefs, which can be true or false; goals, which the agent wants to be true; and plans, which are a collection of actions that the agent performs in order to reach a goal. The author formalizes these notions he gives examples. But it is when he discusses emotions that things get very complicated, for he introduces 78 axioms that he believes are needed for implementing the Ortony-Clore-Collins emotion theory. He does not discuss whether or not these axioms are independent, and from a simplicity standpoint the implementation would be more believable if the number of axioms were a lot less than the ones that are required in the book.
Help other customers find the most helpful reviews
Was this review helpful to you? Yes
No
