Heterogeneous Agent Systems
Heterogeneous Agent Systems
V.S. Subrahmanian, Piero Bonatti, Jûrgen Dix, Thomas Eiter,Sarit Kraus, Fatma Ozcan and Robert RossMIT PressCambridge, Mass.2000xiv + 580 pp.ISBN 0-262-19436-8, hardback£39.95
This very substantial work introduces and develops a new approach to software and computer system development, where an "agent" is a module of code, capable of performing a specific task or set of tasks, and interacting with other agents to achieve some overall result. The first-named author, as well as the two last-named, belong to the University of Maryland and the treatment is associated with a project denoted by the acronym IMPACT, or Interactive Maryland Platform for Agents Collaborating Together. (I find the redundancy of the last two words slightly irritating, but have not been able to think of alternatives that would suit the acronym.) The project is clearly international since the four other authors are respectively from Italy, Germany, Austria and Israel.
The agents are visualised as providing some service that may be called upon by other agents. It is necessary that an agent has an associated interface giving information on the services offered and how they can be invoked. The principle is illustrated with reference to three application areas. The first is to the operation of a department store offering web-based services and dealing with customer enquiries as well as trying to promote business by informing customers about merchandise likely to interest them. The operation is described as depending on a total of five agents, assumed to be implemented in software.
One agent in this example is the Interface Agent, which receives inputs from customers and generates the multimedia output to them. This has two-way communication with a Content Determination Agent, which interacts with a Product Database Agent to produce answers to enquiries as well as promotional material. The Interface Agent also informs a Profiling Agent about the customer input, and this agent prompts the Content Determination Agent about targeted promotional activity. The Profiling Agent takes account of customer credit information which it requests and receives from the Credit Agent.
Of the other two application areas, one is to navigation of an aircraft, with the acronym CFIT, for Controlled Flight Into Terrain. The objective is to integrate and use position data from different sources, each operating through an agent, with the possibility that some sources may mallunction, perhaps due to interference by an enemy or saboteur. The possibility of unreliable agents is a new feature. The third application area is to management of the supply chains for a manufacturing process, where the reliability of suppliers has to be judged and possibly new suppliers requested from a Yellow Pages Agent.
The introduction of a Yellow Pages Agent, and also of a Thesaurus Agent to be called in when agents try to communicate using words not mutually understood,give the scheme an intriguing anthropomorphic character and it is possible to imagine a host of agent types offering a variety of brokerage, consultancy and possibly insurance and even espionage services. These wild speculations (not from the book) should not obscure the fact that this is a serious work in which conventions for agent communication are worked out in detail. Belief logics are invoked, and temporal aspects have to be taken into account, so that, for example, agents involved in the Controlled Flight Into Terrain application do not continue to sound warnings that the flight is heading for a mountain when in fact the planned path has been modified and an appropriate change of direction will occur later but in good time.
One chapter is devoted to Meta Agents, by which is meant agents that are part of the model of the environment on which decision making depends. The example used is a military one, where a variation of the Controlled Flight Into Terrain application is applied to a group of helicopters attempting to destroy a group of tanks on the ground. The tanks can be assumed to be using a CFIT system essentially similar to that of the helicopters, and deductions can be made by simulating this. The authors of the book seem to consider that such modelling is appropriate only in warlike adversarial situations, but in a milder form it could presumably be part of the operation of the Profiling Agent in the department store application. The idea is particularly interesting since it can be argued that a survival advantage of the evolution of consciousness in people and animals is that it allows the use of one's own reactions to an imagined situation to be used as a model of the responses of others with whom there is interaction.
Two potentially important features seem to be missing from the treatment. One is the assignment of priority of access to agents. It is possible to imagine,for example in the department store example, that a forward-planning agent might decide to run a long series of simulations of situations that might arise, in order to compile statistical data. It would be counterproductive if this were allowed to monopolise the Content Determination Agent and other essential agents to the extent that current customer enqutries were delayed. There also appears to be no provision for automatic self-improvement by learning, either of the internal workings of agents or of their interactions.
Both of these features indicate the need for a feedback of "worth" of the kind visualised by Selfridge (1959) in his "Pandemonium" scheme. A generalisation has been introduced by this reviewer as "significance feedback"and will be reviewed in Andrew (2001). Many of the considerations relevant to a multi-agent system presumably apply also to human organisations such as firms. This suggests that the absence of a facility for self-improvement is a serious defect, in view of the insistence, in numerous publications by, notably,Stafford Beer, that the essential channels of communication within a firm can be quite different from what is suggested by a conventional organisation chart.
A large number of AI projects in the past have considered interaction between agents, but the present book introduces and develops this general approach in essentially practical contexts, and will certainly be accepted as essential reading. For a relatively new development, the approach is specified in a surprising amount of detail and it is difficult to belleve that aspects will not be revised as experience is gained. A specific criticism of the presentation is that a lot of acronyms are used without explanation, and the inclusion of a glossary would have been helpful.
Alex M. Andrew
