英文原文：TITLE: MAPCon: AN EXPERT SYSTEM TO CONFIGURECOMMUNICA TIONS NETWORKSCONTACT: H. Van Dyke Parunak, James Kindrick, and Tihamer Toth-FejelIndustrial Technology InstitutePO Box 1485Ann Arbor, MI 48106van%iti@(313) 769-4049TOPIC: Case Study: configuration.ABSTRACT: MAPCon is an expert system that performs off-line parameter configuration for local area networks that use MAP, theManufacturing Automation Protocol. This paper describes theconfiguration task in general and MAPCon in particular, anddescribes its performance as a function of network size.MAPCon: AN EXPERT SYSTEM TOCONFIGURE COMMUNICATIONS NETWORKSH. Van Dyke Parunak, James Kindrick, and Tihamer Toth-FejelIndustrial Technology Institute1. Problem DefinitionThis section outlines the challenge of network management in general and the MAP architecture in particular, and describes the specific functions that MAPCon fulfills.1.1. The Challenge of Network ManagementManagement of multi-vendor networks is difficult, since most network management products are designed for a single product line. Network management is especially critical in manufacturing LANs, where real-time manufacturing operations rely heavily on consistent network operation. One way to address this problem is to conform the network to a standard, such as the Manufacturing Automation Protocol (MAP). MAPcon 11, the system described in this case study, is the second generation (11 of a knowledgebased system designed to assist in configuring MAP networks . This function, the first that must be accomplished when designing a network, requires the correlation of a large number of parameters in complex ways in order for thenetwork to behave properly.1.2. MAP Architecture MAP complies with the architecture of a set of international standards defined by the International Organization for Standardization (ISO), based on a reference model for Open Systems Interconnection (OSI). OS1 decomposes the problem of communicating reliably between applications into seven layers: physical, data-link, network, transport, session, presentation, and application.The physical and data-link layers of MAP use the IEEE 802.4 token busspecifications, in contrast to the CSMA/CD technology familiar from Ethernet. The physical layer can be either broadband or carrierband. WCon distinguishes between broadband and carrierband, and can configure both types of networks. In MAP, the network layer uses the ISO connection-less protocol or the =called Internet protocol. The transport layer uses the IS0 class 4 transport protocol. 1,Session uses ISOsession version 2, which is a small, basic subset of ISO session protocol. Presentation uses the ISO presentation layer.MAP specifies several application layer protocols: a manufacturing message specification protocol; an association control service element; a file transfer, access and management protocol; a directory services protocol; and a network management protocol. The protocols at layers 1, 2, and 7 of MAP were partially or completely designed with manufacturing automation in mind.The entities in the MAP network architecture include stations or end systems, subnetworks or LAN [Local Area Network) segments, and interconnections or intermediate systems.Stations combine hardware and software to provide communications according to the MAP specification. They can have either a full MAP, a mini-MAP, or a MAPIEPA (Enhanced Perjormance Architecture) configuration. The current version of MAPCon handles only full MAP stations.A subnetwork or LAN segment is a section of a local area network on which all stations share the same token. All stations. on a segment can directly cornmunitate to all other stations on the same segment without any intermediate systems. Subnetworks are of two types: broadband and carrierband. MAPCon handles both types of subnetworks.Interconnection entities (intermediate systems) are devices that connect multiple subnetworks to form the overall MAP network. They are of three types: bridges, routers, and gateways. A bridge interconnects two or more subnetworks with similar media access control services. A router interconnects two or more subnetworks or nptworks of different types. A gateway interconnects two or more subnetworks or networks of different network architecture by performing protocol translation. The current version of MAPCon supports bridges and the MAP side of routers and gateways.1.3. Configuration ManagementConfiguration management is the collection of network management activities that allow the user to know and control the arrangement and state of a network and its entities. The present version of MAPCon lets the user set the network configuration off- line, and thus know its arrangement and state. A future version will let the user control the network configuration as the network operates. The following paragraphs discuss the functions currently provided by MAPCon.Adding or deleting stations in the subnetwork defines the actual topology of the network. Stations and interconnections that are attached to a subnetwork are specified in order to configure the network. This capability not only allows for examining the current configuration of the network but also permitsincremental changes to the configuration.Subnetworks are interconnected by attaching interconnection devices such as bridges, routers, and gateways. Adding such devices requires setting or modifying the relationships between two or more subnetworks.Users need to set the initial characteristics of entities and modify them if the original settings are inconsistent. In general, the characteristics that are set or modified correspond to operational parameters and statistical counters required and maintained by the entities. Each station or interconnection has several layers.Depending on the configuration of the station or the interconnection, specific layers must be present according to the MAP 3.0 specification. Several characteristics in each layer must be set properly across the subnetwork and network in order to configure an operational network. These characteristics include operational parameters, timers, and thresholds on statistical counters. Some of these characteristics are derived from network-wide parameters such as the type of the traffic intended on the network and the nature of the environment under which network will operate (low noise, medium noise, and high noise). Other characteristics are derived from one another. For example, inactivity time at transport is derived from retransmission time and number of retransmissions.Stations and interconnections within a subnetwork must have unique names and MAC addresses. However, there are certain restrictions on names and addresses from a global network point of view. It may be possible to have identical MAC addresses in subnetworks interconnected by a router. The configuration functionmust check address consistency.Similar to names and addresses, network managers and load servers must be assigned to stations and interconnections. All stations and interconnections need an associated manager. It is possible to have multiple managers in a network, up to one per subnetwork. A load server for the network is needed if there are anyloadable stations or interconnections in the network configuration.Configuration also requires consistency checks to make sure that all entities are configured correctly. Mechanisms must be provided to recognize and identify the inconsistencies for corrective actions. Consistency in addresses, network managers, and load servers must be checked before the configuration objective is completed.2. Previous ApproachesAI applications in network management have concentrated in the area of diagnosis and fault management rather than configuration.Expert systems for network configuration have been introduced by some vendors, including BBN and CASE Communications. Technical case studies of these commercial products have not been released, so few details of their structure and function are available. At least one uses OPS-83 rules as its major knowledge representation mechanism. While MAPCon does use some OPS rules, it relies heavily on knowledge represented in frames, to support a structural model ofthe network. This type of structure has also been reported in BBN’s Designe t.Previous systems deal with wide-area networks, and focus on problems of routing and resource allocation. One suggested formalization is to allocate network capacity by maximizing an economic criterion related to network revenues. 1151 However, the complexity of the problem precludes use of standard operations research techniques to address this problem. [I21 BBN’s offering includes a facility to allocate parts, such as cables, boards, and racks, to node sites, much along the lines of RI.MAP’S open architecture and bus structure remove problems that other networks must address. Because MAP networks use a bus rather than a point-bpoint architecture, routing is less of a problem. Because MAP is an open architecture, supporting equipment from different vendors, problems of configuring the boards and cables in a single station are also outside of its scope. The major problem addressed by WCon is consistency of parameter settings across different stations. This problem is not a critical one in the networks served by the other programs named above, because they are predominantly single-vendor networks, in which consistency among operating parameters can be enforced at the design or manufacturing stage of a station’s life. The open architecture that allows MAF’Con to push the ”board and cable” problem back to the manufacturer, also forces the network administrator (rather than the manufacturer) to worry about parameter consistency. Furthermore, MAP differs from these other networks in conforming to IS0 protocols for the OS1 seven layer model. The complexity of these protocols and the level of service they provide require many station parameters (62 in MAP 3.0), leading to the need for MAPCon.3. Our ApproachIn this section we discuss some details of MAPCon’s inner structure and function. We detail the different techniques of knowledge representation that it uses, and show how it performs both synthesis and analysis in its reasoning. Then we sketch how the interplay between these reasoning domains will increase as MAF’Con evolves from a configuration tool toward a full-fledged network supervisor.3.1、Knowledge Representation The existing network configuration systems on which details are available use heuristic knowledge in rules as their main knowledge resou rce. MAPCon’s central knowledge structure is a semantic net model of the network being configured, constructed in the Carnegie Representation Language. It does use rules, but as a means of constraining relations among these objects more than to capture shallow information about configuration preferences, It also uses some procedural computation.3.1.1. FramesThe major domain knowledge involved in MAP network configuration is the identity and interrelation of the network entities, so a frame model is both a natural way to store this information and a reasonable basis for propagating constraints among related entities.The features of frames that are most important for MAPCon aremodularity, connectivity, inheritance, and demons.MAPCon interactively guides the user in the construction of the network model. Each network entity (including subnetworks, stations, intermediate systems such as routers or bridges, and points of attachment) corresponds to a CRL frame, and its configurable parameters are slots in the frame.MAPCon frames are related to each other by relations such as my-intermediate-system, relating a point of attachment to its intermediate system, and has-elements, relating subnetworks to their component stations and points of attachment. Frames conveniently permit the system to maintain a deep understanding about the connectivity of the network. For example, intermediate systems can be one of three types. The type of intermediate system connecting two subnetworks determines the quality of service provided by the network across that connection. One type of intermediate system supports address translation, while another type does not. Thus, two interconnected subnetworks may or may not be required to share an address space, depending upon the type of intermediate system providing the connection.MAPCon frames make extensive use of inheritance. For example, a subnetwork frame shares many structural characteristics of a network frame, while the three kinds of intermediate systems have numerous points of similarity. The class structur e of MAF’Con’s ontology permits easy addition of new kinds of entities.In developing a consistent set of parameters, a change in one parameter may have a cascading effect on others. MAPCon uses demons attached to critical slots to propagate these constraints among related objects. For example, mazimum-ring-maintenance-rotation -time is a station parameter based on a user input value for the enclosing subnetwork.The key subnetwork input value propagates to the component stations over the has-elements relation, and constrains the value of the dependent parameter of each station.3.1.2. Rules and Procedural KnowledgeIn classical expert systems, rules are used to capture the heuristic, shallow knowledge of human experts. While =me MAPCon rules serve this function, most constrain relations among slots on the objects representing MAP network elements. Some relations can be constrained through inheritance among objects. In other cases, particularly when several parameters of a single element interact, procedural code attached by demons to affected slots. In the remaining cases, rules permit straightforward declarative specification of interdependencies that would otherwise require complicated, error-prone and hard-to-modify procedural code. For example, validation of an assigned network manager may require searching multiple interconnected subnetworks and is much easier to understand and implement as a (declarative) rule than an extensive (procedural) search through the semantic model.3.2. Synthesis and AnalysisIt is useful to distinguish between two classes of reasoning objectives: synthesis and analysis. Like many other real-world systems, MAPCon does both,switching between them when appropriate.The basic synthesis problem seeks, given a set of elements and a set of constraints among those elements, to assemble from the elements a structure that satisfies the constraints. The term “planning“ is often used loosely to describe this process, though we prefer to reserve it for a more specific case. We distinguish three major types ofsynthesis, which differ in the class of data [I71 used to represent time in its constraints. We can thus define configuration as synthesis in nominal time; planning in the strict sense as synthesis in ordinal time; and scheduling as synthesis in interval time.Analysis begins with a known structure and reasons about the relation between its behavior and the elements that make it up. The two major forms of analysis are prediction, which reasons from the structure and the behavior of its elements to the behavior of the whole, and interpretation, which reasons from the structure and its observed behavior to the state of its elements. Interpretation in turn can involve monitoring to detect unexpected behavior and diagnosis to explain that behavior.MAPCon is primarily a synthetic system, performing static configuration in the nominal time domain. 1181 MAF’Con determines values for 62 interdependent, configurable parameters for each configurable element (station or point of attachment) of the MAP network being modeled.The parameter setting process must follow a partial time ordering, but the resultant configuration itself is nominal with respect to time.4. PerformanceThe performance of an expert system can be measured both in terms of how well and how quickly it executes its task. It is also interesting to record the match between domain and knowledge engineering environment by noting the amount of custom code needed to build an application.4.1. How Well does MAPCon Perform?WCon's task is to determine whether the parameters in the components of a given network can be configured consistently with one another, and if they can, to perform that configuration. Its structure is such that it always succeeds in configuring a configurable network and in properly flagging an unconfigurable one. Thus, it performs its task well.4.2. How Fast does MAPCon Perform?Since expert systems are often applied in domains that suffer from combinatorial complexity, it is important to understand their speed performance as a function of problem size. While analytic complexity bounds may be available for simple cases, the most straightforward way to assess the speed performance of a full-scale system is to gather execution statistics for it. To this end, we have carried out some preliminary experiments on MAPCon.Leaving aside the user interface, the actual configuration task has three computational phases: procedural computation on the parameters supplied by the user;rule-directed reasoning; and a final phase of procedural computation. We recorded CPU time (on a TI Explorer) for each of these phases, and total number of rule invocations, for five test networks that differed from one another both in the number of stations and in the number of subnetworks.Exhibit 1 shows the number of stations and subnetworks in each test network. Note that configurations 2 and 4 both have ten stations, while 3 and 5 both have twenty stations. The differences between them are thus due only to the division into subnetworks in configurations 4 and 5. In both cases, this division takes the form of a star configuration, with a single router connecting the subnetworks.In all five configurations, the total number of rule invocations per configuration is linear in the number of stations, regardless of the number of subnetworks.In the plots accompanying this discussion, we use lower case 'x' and '0' to represent configurations 1, 2, and 3, which we will often call collectively "123" and which compare number of stations in a single network. Upper case 'X' and '0' represent configurations 1, 4, and 5 (collectively '145'), which compare results for different numbers of networks. Configuration 1 appears in both sets.Exhibit 2 plots raw CPU time to execute the OPS rules portion of the inference cycle, as a function of number of stations, for configurations 123. This graph shows a roughly linear increase in execution time with size of working memory, a result in keeping with more general results on the Rete algorithm. But there is a slight convexity (cupped shape) to the curve. To see more detail, we fit by eye a straight line (y = 10.4~ - 25.7) to the data, and subtract it out, leaving the residuals plotted in Exhibit 3. We have removed slope and magnitude information from these residuals, and in exchange can see more clearly the slight (note the difference in scale between Exhibits 2 and 3) convexity hinted at in the earlier exhibit.翻译原文著作（期刊）名称：MAPCon: AN EXPERT SYSTEM TO CONFIGURE COMMUNICATIONS NETWORKS作者：Van Dyke Parunak, H.; Kindrick, J.; Toth-Fejel原文所在位置：IEEE Xplore数据库原文出版时间：2002原文出版地点：美国MAPCon：配置网络的专家系统1、问题定义本节总体概述了网络管理的挑战并且特别说明了MAP体系结构，并介绍了满足MAPCon的特殊功能。