Dynamic Index Construction with Deep Reinforcement Learning

Abstract Thanks to the rapid advances in artificial intelligence, a brand new venue for database performance optimization is through deep neural networks and reinforcement learning paradigm. Alongside long literature this regime, an iconic crucial problem index structure building. For problem, prior works have largely adopted pure learning-based solution replacing traditional methods such as B-tree Hashing. While line of research has drawn much attention field, they ubiquitously abandon semantic guarantees also suffer from loss certain scenarios. In work, we propose Neural Index Search (NIS) framework. The core framework train search policy find near optimal combination plan over existing structures, together with required configuration parameters associated each plan. We argue that compared against approaches, NIS enjoys advantages brought by chosen conventional structures further robustly enhances any singular structure. Extensive empirical results demonstrate our achieves state-of-the-art performances on several benchmarks.