Abstract

Speculative plan execution can be used to significantly improve the performance of information gathering plans. However, its impact is closely tied to the ability to predict data values at runtime. While caching can be used to issue future predictions, such an approach often scales poorly with large data sources and is unable to make intelligent predictions about novel hints, even when there is an obvious relationship between the hint and the predicted value. In this paper, we describe how learning decision trees and transducers can lead to a more efficient value prediction system as well as one capable of making intelligent predictions about new hints. Our initial results validate these claims in the context of the speculative execution of one common type of information gathering plan.

Date

February 6, 2026

Authors

Greg Barish, Craig A Knoblock

Conference

WebDB

Pages

77-82