Optimizing Query Evaluations using Reinforcement Learning for Web Search
This work addresses efficiency improvements for large-scale web search engines like Bing, though it is incremental as it builds on existing match planning methods.
The paper tackled the problem of inefficient candidate generation in web search by formulating match planning as a reinforcement learning task, resulting in up to a 20% reduction in index blocks accessed with minimal quality degradation.
In web search, typically a candidate generation step selects a small set of documents---from collections containing as many as billions of web pages---that are subsequently ranked and pruned before being presented to the user. In Bing, the candidate generation involves scanning the index using statically designed match plans that prescribe sequences of different match criteria and stopping conditions. In this work, we pose match planning as a reinforcement learning task and observe up to 20% reduction in index blocks accessed, with small or no degradation in the quality of the candidate sets.