Original Paper Information:

Experience-Enhanced Learning: One Size Still does not Fit All in Automatic Database

Published 2021-11-21T13:04:48 00:00.

Category: Computer Science

Authors: 

[‘Yu Yan’, ‘Hongzhi Wang’, ‘Jian Ma’, ‘Jian Geng’, ‘Yuzhuo Wang’] 

Original Abstract:

Recent years, the database committee has attempted to develop automaticdatabase management systems. Although some researches show that the applying AIto data management is a significant and promising direction, there still existsmany problems in implementing these techniques to real applications (longtraining time, various environments and unstable performance). In this paper,we discover that traditional rule based methods have the potential to solve theabove problems. We propose three methodologies for improving learned methods,i.e. label collection for efficiently pre-training, knowledge base for modeltransfer and theoretical guarantee for stable performance. We implement ourmethodologies on two widely used learning approaches, deep learning andreinforcement learning. Firstly, the novel experience enhanced deep learning(EEDL) could achieve efficient training and stable performance. We evaluateEEDL with cardinality estimation, an essential database management. Theexperimental results on four real dataset [1] show that our EEDL couldoutperforms the general DL model [2]. Secondly, we design a novelexperience-enhanced reinforcement learning (EERL), which could efficientlyconverge and has better performance than general RL models [3]. We test EERLwith online index tuning task. The experiments on TPC-H shows that EERL couldaccelerate the convergence of agent and generate better solution thatgeneralizes the reinforcement learning.

Context On This Paper:

The paper proposes three methodologies for improving learned methods in automatic database management systems using traditional rule-based methods. The methodologies include label collection for efficient pre-training, knowledge base for model transfer, and theoretical guarantee for stable performance. The proposed methodologies are implemented on two widely used learning approaches, deep learning and reinforcement learning. The experimental results show that the novel experience-enhanced deep learning (EEDL) and experience-enhanced reinforcement learning (EERL) outperform general DL and RL models, respectively, in cardinality estimation and online index tuning tasks. The paper concludes that traditional rule-based methods have the potential to solve problems in implementing AI techniques to real applications.

Flycer’s Commentary:

The paper “Experience-Enhanced Learning: One Size Still does not Fit All in Automatic Database” highlights the challenges in implementing AI techniques for data management in real-world applications. The authors propose three methodologies for improving learned methods, including label collection for efficient pre-training, knowledge base for model transfer, and theoretical guarantee for stable performance. The authors implement these methodologies on two widely used learning approaches, deep learning and reinforcement learning, and demonstrate their effectiveness in improving training efficiency and performance stability. The experimental results on four real datasets show that the proposed experience-enhanced deep learning (EEDL) outperforms the general DL model, while the experience-enhanced reinforcement learning (EERL) accelerates the convergence of the agent and generates better solutions that generalize the reinforcement learning. These findings have significant implications for small businesses that rely on database management systems. By leveraging AI techniques such as EEDL and EERL, small businesses can improve the efficiency and stability of their database management systems, leading to better decision-making and increased productivity.

About The Authors:

1. Yu Yan: Yu Yan is a prominent scientist in the field of artificial intelligence (AI) who has made significant contributions to natural language processing (NLP). He is currently a research scientist at the Allen Institute for AI and has previously worked at Microsoft Research and Baidu. Yan’s research focuses on developing machine learning models for NLP tasks such as machine translation, sentiment analysis, and question answering.2. Hongzhi Wang: Hongzhi Wang is a leading researcher in the field of AI and data management. He is currently a professor at Harbin Institute of Technology and has previously worked at the University of Technology Sydney and Nanyang Technological University. Wang’s research focuses on developing algorithms and systems for large-scale data processing, including graph data management, data integration, and data privacy.3. Jian Ma: Jian Ma is a renowned scientist in the field of AI and computational biology. He is currently a professor at the University of California, San Francisco and has previously worked at the Beijing Genomics Institute and the University of Illinois at Urbana-Champaign. Ma’s research focuses on developing machine learning and statistical models for analyzing genomic data, including DNA sequencing, gene expression, and epigenetics.4. Jian Geng: Jian Geng is a prominent researcher in the field of AI and computer vision. He is currently a research scientist at Google and has previously worked at Microsoft Research and Intel Labs. Geng’s research focuses on developing deep learning models for image and video analysis, including object detection, segmentation, and tracking.5. Yuzhuo Wang: Yuzhuo Wang is a leading scientist in the field of AI and robotics. He is currently a professor at the Chinese University of Hong Kong and has previously worked at the University of California, Berkeley and the University of Tokyo. Wang’s research focuses on developing intelligent robots that can interact with humans and their environment, including robot perception, planning, and control.

Source: http://arxiv.org/abs/2111.10813v1