1986 Multilevel Constraint-Based Configuration Article
0 likes59 views
The document describes MyCon, a prototype expert system designed for configuring computer systems by relieving support engineers from manual configuration tasks. It employs a multilevel constraint-based algorithm to manage the placement of devices on I/O channels, ensuring configurations meet both required and suggested constraints, maximizing performance while minimizing costs. The MyCon interface facilitates user interaction and allows for adjustments, providing flexibility in evaluating and comparing configurations.
1986 Multilevel Constraint-Based Configuration Article
- 1. Multilevel Constraint-Based Configuration The goal of MyCon a prototype expert system for configuring computer systems, is to relieve the support engineer of the tedious task of configuration of a customer order. by Robert I. Marcus A new algorithm for handling computer system conïŹguration problems has been implemented in a prototype called MyCon developed at the Hewlett- Packard Knowledge Systems Laboratory. The original MyCon was written in Portable Standard Lisp using the HP AI Workstation environment1 consisting of an objects package, a frame-based package, and an EMACS- type editor, NMODE. This environment made it possible to construct an interface in which all the knowledge about the devices, constraints, and conïŹgurations of a system is easily accessible to the user. This ïŹrst prototype focused on placing devices on I/O channels. This prototype has been converted into HP standard Pascal and is running on the HP 3000, conïŹguring the HP 3000 Series 68. The latest version also performs slot, junction panel, and cable conïŹguration. The ultimate goal is to produce a program that can free the support engineer from the tedious task of logical conïŹguration of a customer order. The fundamental idea of the MyCon algorithm is very simple, but the algorithm appears to have a wide range of applications. The basic principle is that at every stage of the conïŹguration process there are explicit constraints that cannot be violated. The basic problem solved by MyCon is to construct a system conïŹguration that satisïŹes a series of constraints supplied by an expert. The constraints fall into two differ ent classes: required and suggested. A conïŹguration is feasible only if it satisïŹes all of the required constraints. The suggested constraints are guides for selecting the best choice among the feasible conïŹgurations. The prototype performs the logical conïŹguration of a computer system. This problem consists of taking a set of devices and attaching them to I/O channels to maximize performance at a reasonable cost. Typical required constraints are âThe maximum number of devices that can be attached to a channel is 6â and âHigh-speed devices must be
- 2. attached to high speed channelsâ Typical suggested constraints are âSpread the high- speed devices evenlyâ and âKeep low-speed devices on low-speed channelsâ. The difïŹcult part of the problem is to use the knowledge embedded in the suggested constraints to control and evaluate the system conïŹguration. The method of multilevel constraints was developed to handle this part of the problem. Previous work on conïŹguration problems, such as the Rl system developed by Digital Equipment Corporation have used forward chaining rules to guide the conïŹguration. The advantage of a rule-based or constraint-based system is the ease of maintenance and updating. In this paper, multilevel constraints are discussed as an alternative or supplement to forward chaining. The method can be extended to any problem in which there are qualitative preferences among a range of feasible solutions. Basic Method as Used by MyCon The steps in the basic method are: 1. Choose an initial algorithm that can generate candidate partial conïŹgurations as steps towards a ïŹnal conïŹguration. 2. Use the required constraints as a pruning mechanism to ensure that the initial algorithm will produce only feasible conïŹgurations. 3. Add suggested constraints as necessary to maximize the quality of the ïŹnal conïŹguration and to guide the con ïŹguration away from unsuccessful partial conïŹguration that will require backtracking. 4. If the initial algorithm generates a sufïŹcient number of possibilities, and all the required constraints are stated, and the suggested constraints are chosen well, then the method provides an efïŹcient algorithm for optimal conïŹguration. The idea of multilevel constraints is to assign each constraint to a level determining when it is active and must be satisïŹed. In MyCon, the required constraints are assigned a level of 100. Suggested constraints are assigned lower levels ranging from 0 to 90 in steps of 10. The program begins attaching devices to channels keeping all constraints active. If a feasible conïŹguration cannot be completed, then constraints at level 0 are dropped, the level is incremented by 10, and an attempt is made to attach the remaining devices. This process of dropping constraints is continued until the conïŹguration is successfully completed or the level reaches 100. This method offers several advantages over a strictly for ward chaining approach. The number of constraints is smaller than would be necessary with forward chaining. It is not
- 3. necessary to introduce additional predicates to resolve conïŹicts and to prevent the premature ïŹring of rules. It is possible to evaluate the results of the conïŹguration by keeping track of the suggested constraints it vio lated. Finally, the user of the system can suggest an alternate conïŹguration for the program to evaluate and compare with the original result. This last feature is very important, because it allows a user to determine why the program has selected a particular conïŹguration and which alternatives are equally suitable. The MyCon Prototype The user interface to MyCon is an interactive version of the form used by salespersons when taking a customer order. The form will only accept orders that satisfy all the basic system requirements regarding maximum and mini mum numbers of devices and peripherals. MyCon can also determine if an order contains too many devices to be conïŹgured legally and will advise the user on how to reduce the order. When a correct order has been entered, the program chooses the number of high-speed I/O channels and the system disc following guidelines suggested by an expert in conïŹguration. The user is informed of the reasons governing these choices and is allowed to overrule them. Next, MyCon attempts to place devices on the I/O channels using the active constraints to determine if the placement can be done. The devices are arranged in an order based on heuristics used by experts. In general, devices that are subject to more constraints, such as the system disc, are placed ïŹrst. The user is able to watch the devices being placed one at a time with levels reported if desired. A device is placed on the ïŹrst I/O channel for which no active constraint is violated. If a device can't be placed at a level, then no further attempt is made at that level to place devices of the same type. When all devices possible have been attached at a given level, Mycon checks to see if the conïŹguration is completed. If any devices remain to be placed, the level is raised and constraints below the current level are dropped. If the level reaches 100 without a successful conïŹguration, the user is informed that it is necessary to reduce the number of high-speed I/O channels, which will lower the performance of the system. If the user decides to continue, the number of high-speed I/O channels is reduced, all devices are removed from the channels, the level is set at 0, and the conïŹguration is restarted. This process is repeated until a successful conïŹguration is produced or no further reduction is possible. The last possibility has been eliminated in practice by using a
- 4. maximal packing estimator on the conïŹgurability of the original order before starting conïŹguration. Finally, the user can reconïŹgure the system by adding, moving, and deleting devices. MyCon will evaluate the user's conïŹguration and compare it with the original result. Examples of Multilevel Constraints Multilevel constraints serve a dual purpose in conïŹguration. In one sense they are a means of knowledge representation that is easily understood. The levels convey an estimate of the importance attached to satisfying a constraint. However, the levels are also a control mechanism which can be used to govern the dynamics of the conïŹguration process. For example, to ensure that high-speed devices are spread evenly on high-speed channels, the following constraints and levels are used: Maximum of one device on high-speed channels Level = 10 Maximum of two devices on high-speed channels Level = 20 Maximum of three devices on high-speed channels Level = 30 Maximum of four devices on high-speed channels Level = 40 Maximum of five devices on high-speed channels Level = 50 A fundamental constraint of the system is: Maximum of six devices on high-speed channels Level = 100 The placement of a low-speed device on a high-speed channel at an early stage of the conïŹguration would not strongly affect the performance of a successful conïŹguration. However, it might stop the later placement of a high speed device and prevent the conïŹguration process from succeeding. Therefore, it is important that no low-speed devices be placed on high-speed channels until all high speed devices have been attached. Since the constraints above may prevent this until the level is greater than 50, the following constraint is enforced: No low-speed devices on high-speed channels Level = 60 It is possible to have MyCon maintain its knowledge base by insisting that certain inequality relations between constraint levels be enforced. For example: Level of (No low-speed devices on high-speed channels) > Level of (Maximum of five devices on high-speed channels). The constraints can be chosen to avoid having devices placed on the same channel. For example:
- 5. Minimum of (Number of system discs attached to channel) AND (Number of magnetic tape drives attached to channel) = 0. Level = 90 Number of high-speed master discs attached to channel < = 1 Level = 30 It is also possible to insist that certain devices not be placed with those of another type. For example: (Number of slave discs attached to channel = 0) OR (Number of master d i s c s a t t a c h e d t o c h a n n e l) = 0 . L e v e l = 1 0 0 Observations It has been possible to translate all of the required and suggested constraints for device placement into the form of multilevel constraints easily. The selection of levels for the suggested constraints required some analysis and experimentation. The easy-to- understand form of the constraints facilitates addition, deletion, modiïŹcation, and testing by a conïŹguration expert untrained in program ming. The expansion of the prototype to handle placement into slots and junction panels greatly increased the knowledge required by MyCon. This was accommodated by adding constraints to the device placement phase of the program. Using these additional constraints made possible a procedural implementation of the slot and junction panel con ïŹguration. The use of constraints also permitted an easy transformation of the program from Lisp to Pascal. Using Lisp as a development language hastened development of the prototype by providing ïŹexibility in choosing data structures and an excellent debugging environment. The conversion into Pascal was done for reasons of portability and maintainability. Conclusion The method of multilevel constraints has proven very effective for logical conïŹguration of computer systems. It combines simplicity and power in knowledge representation and control. The method can be extended to many problems in which a system is constructed systematically using knowledge of a domain. It can be used as a supplement or alternative to procedural or forward chaining methods. The constraints complement these methods by telling the program what actions not to take. Constraints seem to be a natural way for experts to ex press their knowledge of a domain. By tracing the constraints and the levels it is possible to monitor very closely the progress of the conïŹguration and detect gaps or errors in the knowledge.
- 6. The real power of the algorithm has not been fully tested by the logical conïŹguration problem. For example, express ing multilevel constraints in mathematical form allows the introduction into expert systems of the tremendous amount of software constructed for linear and mathematical programming. In the ïŹeld of artiïŹcial intelligence, there has been extensive work on constraint satisfaction and propagation; this could be used to extend the present algorithm. In the future, it seems possible that the method of multilevel constraints could become a standard tool of knowledge engineering. Acknowledgments Keith Harrison introduced me to the conïŹguration problem and suggested many improvements in the user interface. References 1. M.R. Cagan, "An Introduction to Hewlett-Packard's AI Workstation March Hewlett-Packard Journal, Vol. 37, no. 3, March 1986. 2. R.M. Stallman, EMACS ĂąâŹâ The Extensible, Customizable, Self- Documenting Display Editor, Memo No. 519, ArtiïŹcial Intelligence Laboratory, Massachusetts Institute of Technology, June 1979. 3. J. Sys "R-l: A Rule-Based ConïŹgurer of Computer Systems," ArtiïŹcial Intelligence, Vol. 11, no. 1, 1982.