MS Thesis: Understanding AI Patent Search: Discrete Action Space Exploration for Boolean Query Generation
May 2026
My Master's Thesis explored ways to explain the results of AI Patent search tools that are opaque, a problem for inventors and patent examiners. I created a system to generate boolean queries that would reproduce the results of an AI-based patent search tool, and iteratively improve the query using machine learning and search tools to explore improvements to the query.
Abstract
Patents play a vital role in protecting innovation and shaping economic growth. Inventors and patent examiners rely on effectively searching existing patents to determine whether proposed patents are novel, or already covered by prior art. AI-based patent search methods can offer utility in this process, but operate as black-boxes and lack the transparency and explainability required in a legal setting. We present an approach to dynamically refine explainable boolean queries which recreate the results of the AI-based search. To reduce the complexity of optimization on such a large search space of possible candidate terms we use ranking and filtering to narrow the broad search space into a discrete action space of candidate terms. We apply three optimization techniques to the resulting action space: Reinforcement Learning (with PPO), greedy search, and Monte-Carlo Tree Search. We find that greedy search was the most effective in our context.