A Multi-Agent System Simulating Crew Interaction in a Military Aircraft. To appear

  • 格式:pdf
  • 大小:31.05 KB
  • 文档页数:4

A Multi-Agent System Simulating Crew Interactionin a Military Aircraft

Denis Gagné, Jocelyn Desbiens and Georges NaultGroupe de recherche en intelligence artificielle distribuée (GRIAD)Collège Militaire Royal de Saint-JeanComputer Sciences DepartmentRichelain, Québec, Canada, J0J 1R0

AbstractThe development of expert systems using parallel machines is a practice in its infancy, but its future is very promising.Researchers at the Artificial Intelligence Laboratory of the C.M.R. de Saint-Jean have been able to implement a cooperativesystem in which relatively independent agents cooperate to solve a specific problem. The system presented is implemented on adistributed memory parallel machine, enabling several autonomous expert systems executing simultaneously, to work together.This approach is part of a strategy to solve complex problems via a multi-agent distributed system.The problem dealt with--to simulate crew interaction in an anti-submarine aircraft--enabled us to demonstrate the manyadvantages to be gained, in a large number of applications, from the presented choice of software and hardware architecture.The long-term objective of this project is to provide a more productive and economical way of training crews of this aircraft.

Introduction

After years of enthusiasm followed by disenchantment, few areas of research in artificial intelligence still enable us to makeprogress without limitations; most techniques used have shown proof of intrinsic limitations or major weaknesses. Thus todate, only expert systems represent a real, functional success in limited applications. However, attempts to generalize theirexpertise or widen their areas of application have been, for the most part, doomed to failure.

Given the ambitions of AI systems, it is foreseeable that an increase in performance of over 10,000 times what is possible todaywill be required [Wah 88]. We already know that parallelizing the Rete algorithm (at the heart of most expert system shells)can bring about, at most, a 10 to 1 gain in performance [Gup 86]. Several parallel implementations of this algorithm haveshown that the possibility of using specific processors is limited. Theoretical analyses have shown that considerable improvedperformance can only be achieved in certain very specific cases [Bah 92].

Attempts at parallelization via the distribution of rules on separate processors have not been much more promising. In fact,where dependency relationships and intelligent distribution algorithms must be defined, it has been shown that even the bestapproach at distributing rules is at least a quadratic assignation problem (proven NP-complete) [Mol 89].

Our Choice of Paradigm

It appears that it will often be easier and more productive to execute a parallelization of the application or data (when this ispossible) rather than attempt to parallelize its resolution at the algorithmic level. It is therefore always worthwhile to exploreall parallelization possibilities at the application level even before continuing on to later stages.

It is further evident that an appreciable increase in the speed of execution will not be obtain unless we judiciously provide forfactors between two contradictory optimization options; the need to centralize tasks to minimize communications, and the needto distribute tasks as much as possible to increase the level of parallelism attained.

A known fact of system analysis is that most complex systems naturally tend to break themselves into relatively independentsubsystems. Within these subsystems, interaction between elements are frequent and rapid, while interaction between onesubsystem and another are comparatively rare and slow [Cou 77].

As a result of these and several other observations, we have chosen an approach which combines the strong points of twowidely used and proven paradigms, while making sure to avoid their disadvantages. An appreciable gain in performance,attributable to parallelism, is added to the power of expert systems applied to limited areas (in which they excel).

The thesis here is that most complex problems can better be solved via a pendemonium of micro agents, each of themspecializing in a different and very narrow aspect of the problem.

83