Optimizing Operations on B-Trees Using Proximal Policy Optimization and Hierarchical Attention-Based Models

ICLR 2026 Conference Submission17705 Authors

Keywords: Reinforcement Learning, Optimization, Databases, Attention, B-Trees

TL;DR: A hierarchical attention-based model is presented with the goal of optimizing operation scheduling on dynamically managed B-trees.

Abstract: Modern database management systems often rely on B-trees to achieve indexing in an efficient manner. If stored on slow permanent storage devices, write and read operations can become a significant performance factor, as transactional databases require regular additions and deletions. We propose to use a reinforcement learning setup to optimize the write performance of deletes and inserts by aggregating them and optimizing their order of execution. This achieves the goal of minimizing write times during tree updates. We present a small hierarchical attention-based model to parse the content of the tree efficiently. The new architecture allows for level-wise parallel computation and includes caching to improve the inference speed. Our evaluation verifies the applicability and the potential of the proposed framework. We show that we can efficiently compute an embedding in a hierarchical dataset and that the embedding can be used to achieve noticeable performance improvements in B-tree operation scheduling in comparison to accepting operations in their order of arrival.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 17705