SURPRISE '97 - Intelligent Agent Technology in Travel Reservation Systems

Imperial College of Science Technology and Medicine Departments of Computing and Electronic Engineering

May 27 1997

An Agent-Based View of Minds and Intelligence
Nick Moraitakis ( nm1@doc.ic.ac.uk )

Abstract:
It has become a trend in the past few years to view agent systems as expert problem-solving entities, and it is commercialization that has led the agent-based community into seeing agents through application. We cannot fail to notice the similarity of the above observation to the tendency of recent AI research to focus into systems with purpose-built experise and problem-solving capability. But is this a wise path to follow? This article's purpose is to look back and bring up a different set of ideas about agents and intelligence, one that is interested not in systems that 'do things that an intelligent system would do' but in minds and an agent-oriented theory of how they work. Then we will get back to how this broader view of agent systems can be applied by analogy to the design and implementation of large gent applications.







1 A theory of minds

What do we mean by mind and intelligence ? Of course we cannot, and we will not attempt to, provide a definition for these two inter-related notions over which researchers, philosophers and scientists have been endlessly disputing since the question was first asked. (whenever that was)

Alternatively, we can discuss and investigate the ways minds work in order to achieve, not intelligence, but the ability to accomplish tasks which we consider needind intelligence to deal with. This we will do by considering a model proposed by Marvin Minsky in his book [1] The Society of Mind . We'll question on the mechanisms relied upon to make this mind-model capable of exhibiting intelligent behaviour.

The ideas we'll use to understand and investigate this model are not new. Minsky presented those ideas a little more than a decade ago. What is new in our investigations is that, by looking at those ideas, we will try to build insights on a model that might be suitable for designing large agent applications. After all, as we shall see shortly, Minsky's mind theory describes a magnificently complicated in physique, but simple in principle agent-system.



2 Is Deep Blue intelligent?

Recently, a chess-playing computer system called Deep Blue managed to beat humanity's world champion, G. Kasparov . Certainly, beating Kasparov needs great knowledge, skill and experience with chess. But would you say that Deep Blue is intelligent? Or that it is knowledgeable? Or skillful? Deep Blue is a system that "knows" how to play good chess. It does not understand chess. It just plays good chess. And there is a major difference between the two. Probably it is the basis on which we differentiate between expert systems (systems that are good at solving problems on a specific domain) and intelligent systems. (systems that perceive, understand, learn and think in ways similar to human minds)



3 A society of agents

In contrast to experts systems, Minsky proposes a model of a mind that consists of a network of agents that perform tasks so that the interoperation and coordination of tasks within the network leads to problem solving. The agents are interconnected in hierarchies and hetererarchies so that mental tasks consist of decomposing them into smaller tasks and then assignment to agents capable of perfoming them. In the words of the author:


I'll call society of mind this scheme under which each mind is made of many smaller processes. These, we'll call agents. Each mental agent by itself can do some simple thing that needs no mind or thought at all. Yet, when we join these agents in societies - in certain very special ways - this leads to true intelligence.

There are many issues associated with this model. For example we might ask how these agents work, or how they communicate, how they are created and replaced, how they resolve conflicts and disagreements, or how they decide on their course of action... Many more questions can be added to this list. The interesting point though, is that we can identify in it all the issues associated with the design and theory of modern agent-based systems. In expert systems and problem-specific agent systems we usually expect these questions to be answered by our design in some such that the the system will exhibit some pre-defined behaviours. The marvel of the above mind theory though, lies in the fact that what we care about when considering minds is not the performance of predefined behaviours, but the ability to develop behaviours. This ability comes from the way the agents are linked and assigned into dynamic coordination schemes. So non-intelligent components, construct intelligent minds.(or maybe computer systems)



4 Minds and systems

The interesting idea behind this mind theory is its potential for application in general agent systems. If we can think of a human like mind as a vast network of agents working independently and in cooperation (or conflict) to achieve tasks, then why not think of a large agent-based computer network as a structure resembling that of a mind, with agents that perform tasks both to achieve local and global objectives? These agents would cooperate and compete for the resources of the network, while they would also control on global-level the behaviour of the network. Telecommunications companies would be dreaming of networks like this one. Information flow in the world-wide community would become a totally different idea. The mere notions of information , resource and task would have to be given new thought and maybe new meanings.



5 Motivations

Why should we want to build such systems? Mainly because the exploitation of the ingenious mechanisms with wich minds work will yield great benefits in computational power.

Problem solving will be more efficient due to the modularity of processes. Each agent knows well how to do something and contributes with its capability to the performance of a global objective. Agents, thus, have scope and are also dispensable beyond that scope, as soon as their function is obsolete (another agent has been developped for it) or useless.(the task they were part of is no longer performed or their hierarchical predecessor is ceased)

Redundancy (i.e. the existence of more than one agents to do a particular job) in such a sytem will make such systems reliable . It must be noted though that the problem solving mechanisms are more complicated when there exist multiple agents capable of performing a certain task.

Sharing of resources in such systems takes a whole new meaning, as we can consider all resources of the system (storage, agent abilities etc) as system memory and experience. This set of experiences, facts and states that the system owns will become the embodiment of the system's awareness. May this awareness be awareness of its own existence, environment, or whatever is appropriate for the application considered.

If a local objective attempts to solve a problem for which the solution has already been found once (a similar objective has been accomplished earlier) then this experience can be used to solve the new problem too. Experience is inherent in the system. This of course means that the system must have some mechanism to identify similar problems by analogy, or by rule-based methods, and possibly revise problem-solving methods and problem-solving rules.

These are just a few thoughts on how such systems would operate in better way than today's sytems can. Of course they are just ideas and motivations for the development of large-scale agent-based intelligent systems, and the theory and implementation behind them involves considering very fine details on how such mind-like systems work.



6 More questions

Visions like the one we developped in the previous sections might look attractive and challenging, but before delving into their implementation possibilities, we must deal with a new wave of questions arising from their mere conception. Are these systems going to have some centralized hierarchy of agents? If yes, how will they be coordinated? If no, will the goals and objectives they perform be global? Will they be just dynamically determined compositions of local goals? And if the goals are going to be dynamically determined, does this mean that these systems will be able to plan? And if they plan, how will they do this? Does this mean that systems of this kind will have intentions or judgement upon them? Or will they develop awareness, or consiousness of their existence? In Minsky's model minds are composed of unintelligent agents into a single and utter unity that is intelligent. Will large scale systems of agents in apropriate interconnections and operating mechanisms assume the properties of a mental entity on global perspective?

Behind each of these questions lie even mor questions about minds, intelligents, and intelligent systems. It is not the purpose of this article to consider them separately or answer them at all. Instead, what this article tried to do is to capture the idea behind a theory of minds and propose these questions as a starting point for further thought on intelligence and its potential presence in agent systems.



List of References:



---




1	Society of Minds
	Marvin Minsky,
	Usefulness :                     (10)
	Readability :                     (10)
	Content : An investigation of minds, intelligence and how intelligence can be built by lifeless, unintelligent parts.
	On-Line :
	Download Postscript :
	Obtained From : Out of print in U.K. -- I.C. Central Library has a copy
	Notes :




---





2	HAL's Legacy
	Edited by David G. Stork , 1997
	MIT Press
	Usefulness :                     (5 )
	Readability :                     (9 )
	Content : A collection of articles by renowned AI researchers and scientists concerning 2001's HAL supercomputer as related to today's state of the art technological capabilities in AI systems
	On-Line :
	Download Postscript :
	Obtained From : Bookstores
	Notes :




---