Towards full automation of the discovery of heuristics in a nuclear engineering project: Integration with a neural information language

Abstract

FUELCON is an expert system in nuclear engineering. The practitioner of nuclear fuel management uses it to generate configurations of reactor core refueling. The domain expert, on the other hand, employs the system to discover new heuristics for the previous task, based on performance during previous iterations in the same session. Expert use involves a manual phase of revising the ruleset. We expose the building blocks of the design of a new version, that incorporates a neural component to carry out the revision. It uses the system's previous performance it for adaptation and learning better rules. for adaptation and learning better rules. The neural component is based on a particular language and schema for symbolic to recurrent-analog conversion, called NEL, and on reinforcement learning for the adaptation.