Adaptive capture of expert knowledge

Technical Report · Mon May 01 04:00:00 EDT 1995

DOI:https://doi.org/10.2172/93561· OSTI ID:93561

Barrett, C L; Jones, R D ^[1]; Hand, Un Kyong ^[1]

---

1. Los Alamos National Lab., NM (United States)

A method is introduced that can directly acquire knowledge-engineered, rule-based logic in an adaptive network. This adaptive representation of the rule system can then replace the rule system in simulated intelligent agents and thereby permit further performance-based adaptation of the rule system. The approach described provides both weight-fitting network adaptation and potentially powerful rule mutation and selection mechanisms. Nonlinear terms are generated implicitly in the mutation process through the emergent interaction of multiple linear terms. By this method it is possible to acquire nonlinear relations that exist in the training data without addition of hidden layers or imposition of explicit nonlinear terms in the network. We smoothed and captured a set of expert rules with an adaptive network. The motivation for this was to (1) realize a speed advantage over traditional rule-based simulations; (2) have variability in the intelligent objects not possible by rule-based systems but provided by adaptive systems: and (3) maintain the understandability of rule-based simulations. A set of binary rules was smoothed and converted into a simple set of arithmetic statements, where continuous, non-binary rules are permitted. A neural network, called the expert network, was developed to capture this rule set, which it was able to do with zero error. The expert network is also capable of learning a nonmonotonic term without a hidden layer. The trained network in feedforward operation is fast running, compact, and traceable to the rule base.

• Research Organization: Los Alamos National Lab., NM (United States)
• Sponsoring Organization: USDOE, Washington, DC (United States)
• DOE Contract Number: W-7405-ENG-36
• OSTI ID: 93561
• Report Number(s): LA-UR--95-1391; ON: DE95012054
• Country of Publication: United States
• Language: English

Related Subjects

45 MILITARY TECHNOLOGY, WEAPONRY, AND NATIONAL DEFENSE
99 GENERAL AND MISCELLANEOUS
ARTIFICIAL INTELLIGENCE
BEHAVIOR
EXPERT SYSTEMS
KNOWLEDGE BASE
NEURAL NETWORKS
SIMULATION
WARFARE