| 贺文凯* ** ***,支天*,胡杏*,张曦珊* ***,张蕊* ***,杜子东*,郭崎*.基于真值表的函数自动生成的神经网络模型[J].高技术通讯(中文),2024,34(3):265~274 |
| 基于真值表的函数自动生成的神经网络模型 |
| Automatically generating function expressions based on truth tables with neural networks |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2024. 03. 005 |
| 中文关键词: 真值表； 神经网络； 序列模型； 示例编程； 程序综合 |
| 英文关键词: truth table, neural network, sequential model, programming by examples, program synthesis |
| 基金项目: |
| 作者 | 单位 | | 贺文凯* ** *** | （*中国科学院计算技术研究所处理器芯片全国重点实验室北京 100190）
（** 中国科学院大学北京 100049）
（*** 中科寒武纪科技股份有限公司北京 100191） | | 支天* | | | 胡杏* | | | 张曦珊* *** | | | 张蕊* *** | | | 杜子东* | | | 郭崎* | |
|
| 摘要点击次数: 212 |
| 全文下载次数: 139 |
| 中文摘要: |
| 作为目前最常见的程序综合问题，示例编程通过用户提供的输入/输出示例生成程序，为编程能力不足的开发者提供了便利。近年来，示例编程已经被应用于Microsoft Office Excel办公软件的自动编程，以及勘探、测井、航空航天等领域。鉴于目前示例编程鲜有关于二进制流的研究，本文针对基于真值表函数自动生成问题具有函数表达式的语法符号序列中各语法符号的关系与它们的距离大小无关、函数表达式的生成语义规则与布尔向量函数采样的结果长度无关的特点，设计了一种神经网络模型和算法，在程序综合、功能等价和序列匹配的指标上分别取得了70.56%、64.66%、0.6355的结果，分别优于现有最先进的程序综合模型55.07%、49.70%、0.5690。 |
| 英文摘要: |
| Programming by examples is a common program synthesis problem that involves generating programs based on input/output examples provided by users, it offers convenience for novice programmers. Recently, programming by examples has been applied to automatic programming of Microsoft Office Excel, as well as in exploration, logging, and aerospace. In order to address the gap resulting from limited research on binary data flow, the problem of generating function expressions based on truth tables is introduced. This problem has the characteristic that the relationship between the syntactic symbols in the sequence of symbols in the function expression is independent of their distances. Additionally, the generation of semantic rules for the function expression is unrelated to the resulting length of the Boolean vector function sampling. Based on the aforementioned characteristics, this paper introduces a neural network model and algorithm that achieve results of 70.56%, 64.66%, and 0.6355 in program synthesis, functional equivalence, and sequence matching, respectively. These results outperform the existing state-of-the-art program synthesis model, which achieves 55.07%, 49.70%, and 0.569 0, respectively. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
