CRINN: Reinforcement Learning Optimizes Approximate Nearest Neighbor Search
A novel approach to approximate nearest-neighbor search (ANNS) has been developed, utilizing reinforcement learning to create quicker implementations while preserving accuracy. This technique, called CRINN, frames the optimization of ANNS as a reinforcement learning challenge, where the speed of execution acts as the reward. As a result, the system can enhance its performance iteratively, yielding faster ANNS solutions without compromising accuracy. The method was tested on six prominent NNS benchmark datasets, achieving top results on three (GIST-960-Euclidean, MNIST-784-Euclidean, and GloVe-25-angular) and tying for the best on two additional datasets (SIFT-128-Euclidean and GloVe-25-angular). These findings highlight CRINN's superiority over leading open-source ANNS algorithms, with significant implications for AI applications that depend on efficient nearest-neighbor searches, such as retrieval-augmented generation (RAG) and agent-based LLM applications. The research can be found on arXiv with the identifier 2508.02091.
Key facts
- CRINN is a new paradigm for ANNS algorithms.
- CRINN treats ANNS optimization as a reinforcement learning problem.
- Execution speed serves as the reward signal in CRINN.
- CRINN automatically generates progressively faster ANNS implementations.
- CRINN maintains accuracy constraints.
- CRINN was evaluated on six NNS benchmark datasets.
- CRINN achieved best performance on GIST-960-Euclidean, MNIST-784-Euclidean, and GloVe-25-angular.
- CRINN tied for first place on SIFT-128-Euclidean and GloVe-25-angular.
- The paper is available on arXiv with identifier 2508.02091.
Entities
—