TL;DR
This paper introduces R1, a rule-based expert system designed to configure complex computer systems (specifically DEC's VAX-11/780). Unlike previous systems, R1 uses a "Match" problem-solving strategy that avoids generating incorrect hypotheses, leading to a more efficient and direct path to a correct solution. R1 has been successfully deployed in a real-world manufacturing setting, demonstrating the practical viability of expert systems for complex configuration tasks.
What Problem Does This Paper Address?
Configuring complex computer systems, like the VAX-11/780, is a challenging task. These systems involve numerous components (around 90 in a typical VAX system), each with specific connection requirements and constraints. Manually configuring these systems is time-consuming and prone to errors. Existing AI approaches often relied on "Generate-and-Test" methods, which could explore many incorrect configurations before finding a valid one. The paper addresses the need for an automated system that can reliably and efficiently configure these complex systems, mimicking the expertise of human engineers.
Methods
R1 is implemented using OPS4, a production system language. The core of R1's problem-solving lies in its "Match" method. This method is analogous to heuristic search techniques like Hill Climbing or Means-Ends Analysis but is more powerful because the conditions associated with each state are sufficient to guarantee that if a state transition is permissible, the new state will be on a solution path. This means R1 avoids generating false paths and thus never needs to backtrack.
R1's knowledge is stored in three memories:
- Production Memory: Contains R1's permanent knowledge in the form of production rules (if-then statements). R1 has 772 rules, with 480 being domain-specific configuration rules.
- Working Memory: Holds data elements representing the current state of the configuration process, including partial configurations and intermediate results.
- Data Base: Stores descriptions of the components that can be used in a VAX system (420 components are supported), including their relevant attribute-value pairs.
When configuring, R1 retrieves component descriptions, builds partial configurations in working memory, and uses its rules to identify the next valid step, directly moving towards a solution.
Datasets
Not discussed in the paper. The paper focuses on the system's architecture and problem-solving approach rather than specific training datasets. The "dataset" in this context is the customer's order for a VAX system and the internal knowledge base of component descriptions.
Key Insights and Ideas
- "Match" as a superior problem-solving strategy: The paper highlights that by ensuring each rule's conditions lead directly to a valid next step, R1 avoids the inefficiency of exploring incorrect paths, a common issue with "Generate-and-Test" methods.
- Expertise captured in rules: R1 demonstrates that complex domain knowledge, like that of VAX system configuration, can be effectively encoded into a large set of production rules.
- Real-world deployment is achievable: The successful integration of R1 into Digital Equipment Corporation's manufacturing organization signifies a major step in applying AI to practical, industrial problems.
- The importance of detailed component knowledge: R1's data base, containing detailed attribute-value pairs for each component, is crucial for its ability to make informed configuration decisions.
Key Results
- R1 was successfully integrated into DEC's manufacturing organization and is used to configure VAX-11/780 systems.
- In a formal validation procedure, R1 configured 50 orders, with only 12 pieces of errorful knowledge uncovered and subsequently corrected.
- For each of these 50 orders (averaging 90 components), R1 fired an average of 1056 rules and took approximately 2.5 minutes of CPU time.
- Since January 1980, R1 has configured over 500 orders and is also being used by DEC's sales organization.
- R1's output provides more detailed information to technicians than human-generated configurations.
Limitations and Future Work
- Limited domain: R1's expertise is currently confined to VAX-11/780 systems.
- Future work: The authors plan to:
- Augment R1's knowledge to configure other DEC computer systems.
- Enhance R1 to assist with scheduling system delivery dates.
- Develop an interactive version of R1 to provide real-time assistance to customers or salespeople, allowing them to specify system capabilities and have R1 select components.
- Ultimately, create a "salesperson's assistant" that helps customers identify the system best suited to their needs.
Summary updated