![]() |
|||
![]()
|
![]() |
![]() |
![]() |
3.5. DATA MINING AGENTSData mining agents process large datasets to extract useful information. There are two classes of data mining agents. One class uses deductive techniques to mine databases. For example, consider a data warehouse containing performance data on multiple product lines in a manufacturing company. A deductive data mining agent can be built to monitor specified quality metrics, such as the number of product defects. When the agent detects a defect rate higher than a set limit, it performs additional analysis to determine possible causes for the high defect rate. Therefore, the agent automates the deductive reasoning of business analysts responsible for manufacturing quality. The power of such an agent is that it exhaustively analyzes manufacturing quality on a continuous basis and alerts supervisory management of slippage in quality in real-time.
The second class of data mining agents applies inductive classification techniques such as neural networks and decision trees to large data sets to automatically recognize significant patterns. For example, consider a telecommunications network consisting of thousands of devices carrying huge volumes of voice and data traffic. A device failure can cost the company dearly, not to speak of the inconvenience to customers. By building classifiers based on historical network data at the time of a device failure, data mining agents are built to predict failure ahead of time. 4. SUPPORTING TECHNOLOGIES AND RESEARCH ISSUESTechnologies and research areas supporting agent applications include agent communication languages, mobile and autonomous agent technologies, and work in cooperation and coordination and in learning. 4.1. AGENT COMMUNICATION LANGUAGESWhen agents interact with each other, how do they communicate? The essential motivation is to develop standard language definitions for inter-agent communication in order to enable agents developed in a decentralized fashion to collaborate. For example, an agent shopping for the best interest rates on a housing loan should be able to communicate with the online loan agents from different banks. Clearly, such interaction requires a standard language that all these agents understand. KQML is currently the best-known agent communication language and has been used in prototype applications requiring agents to communicate with each other. KQML highlights several key requirements of a successful agent communication language. It separates the communication language, meant to express communication requests, from the content language, meant to represent domain knowledge. This separation is achieved by a three-layer language model. The message layer and the communication layer are responsible for performing communicative acts, while the content layer carries the domain-specific representations. KQML leaves open the specification of the content language. Therefore, applications are free to select any content language to use with KQML. The set of communication primitives provided by KQML is quite comprehensive and meets the requirements of a variety of application scenarios. KQML can be used easily with common transport mechanisms such as TCP/IP, SMTP, HTTP, and CORBA. The infrastructure for agent communications is evolving rapidly. Object Request Brokers (ORBs) allow object interoperability. ORBs based around the Object Management Group's (OMG) CORBA 2.0 support the Internet Inter-ORB Protocol (IIOP). IIOP is a distributed object inter-process communication (IPC) protocol. IIOP will allow seamless object method access across heterogeneous ORBs, operating systems, and hardware. Three examples of this infrastructure evolution will be described to illustrate what is becoming available to support intelligent agent communication. APM Ltd., of Cambridge, U.K., is pursuing a collaborative program called the Advanced Network System Architecture (ANSA) that uses IIOP. They are working on an IIOP-to-HTTP gateway that allows CORBA clients to access Web resources. Their infrastructure also contains an HTTP to IIOP gateway that will allow Web clients to access CORBA resources. Figure 3 from APM illustrates this infrastructure. Clients will be able to invoke CORBA services, and the ORB (Object Request Broker) will locate and invoke the right service. If the service is outside CORBA, access takes place via the IIOP-to-HTTP gateway (which also converts from IIOP to other protocols such as FTP). The remote object looks like a CORBA object to the CORBA client (URLs look like CORBA objects). Clients outside the CORBA world see CORBA objects as Web objects and will be able to direct HTTP requests to these objects through the HTTP-to-IIOP gateway. Lotus' Domino Web server allows Lotus Notes applications to be viewable as HTML. The server has a CORBA ORB with IIOP. With this server, the gateway between Web resources and CORBA objects is similar to the APM infrastructure.
Netscape Communications Corporation has licensed an IIOP-compliant ORB from Visigenic Software, Inc. for their browser and server software. The Netscape IIOP infrastructure can support distribute method invocation for Java classes, simplifying the development of Web-based distributed applications. Another aspect of agent communications is the need to interface with existing legacy systems and databases. As discussed in Section 3.4 for the C2 software agent, this infrastructure programming can consume a large portion of the software development process when building intelligent agents. Fortunately, the existence of de facto standards, such as ODBC (Open DataBase Connectivity), have simplified some of this interfacing. The effort involved in this aspect of agent communications should not be underestimated, though, especially if high transaction performance is required. One of the lessons learned from the development and deployment of expert systems in the 1980s was the need to build systems on conventional hardware platforms with robust interfaces to existing data and legacy systems.
|
![]() |
|
Use of this site is subject certain Terms & Conditions. Copyright (c) 1996-1999 EarthWeb, Inc.. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of EarthWeb is prohibited. Please read our privacy policy for details. |