探討程式執行對程式合成的影響

Abstract

設計和實現能夠模擬機器行為的程式是人工智慧研究的核心任務之一。傳統的自然語言處理（NLP）方法在生成可讀性高的人類文本方面表現優異，然而在程式執行的準確性和一致性方面通常存在顯著差距。相較之下，神經程式合成（NPS）方法能夠通過直接從規範或示例生成可執行的程式，從而更深入地理解和操作機器行為。傳統自然語言處理（NLP）與神經程式合成（NPS）之間的主要差異在於，後者能夠生成可執行的程式碼，這些程式碼可以執行並檢查其行為。這種能力不僅有助於驗證合成的解決方案，還為從執行軌跡中學習開闢了新的途徑。 在本論文中，我們探索神經程式合成（NPS）與程式執行之間的共生關係，突顯程式執行如何成為合成程式的核心要素。本論文包含兩個主要項目：第一個項目探討學習程式執行或合成是否有助於學習另一個方面。通過分析這兩個方面之間的相互作用，我們研究神經網路如何有效地理解並生成可執行程式碼。第二個項目專注於提升從多樣化示範影片中合成程式行為的效率和有效性，同時提出指標驗證其在實際情境中的應用性。透過結構分析、前例分析和實驗驗證，本論文旨在深化對「程式執行對神經程式合成影響」的理解，並探討程式表示和行為之間的複雜性，以促進創造更安全、更強大的人工智慧系統。

Understanding and synthesizing programs that represent machine behavior is a fundamental aspect of artificial intelligence research. Traditional approaches to natural language processing (NLP) have excelled in generating human-readable text but have largely remained detached from program execution. Conversely, Neural Program Synthesis (NPS) offers a paradigm shift by generating executable programs directly from specifications or examples, thereby enabling a deeper understanding and manipulation of machine behavior. The key difference between traditional NLP and NPS lies in the ability of the latter to generate executable code, which can be executed to examine its behavior. This capability not only facilitates the verification of synthesized solutions but also opens avenues for learning from execution traces. In this thesis, we explore the symbiotic relationship between neural program synthesis (NPS) and program execution, highlighting how program execution serves as a fundamental aspect aiding in the synthesis of programs. This thesis comprises two main projects. The first project investigates whether learning program execution or synthesis aids in learning the other. By analyzing the interplay between these two aspects, we aim to uncover insights into how neural networks can effectively comprehend and generate executable code. The second project focuses on learning to synthesize programs from entire execution traces. We aim to enhance the efficiency and effectiveness of neural program synthesis from diverse demonstration videos and propose realistic metrics to examine whether the method can be applied in more practical scenarios. Through a combination of structural analysis, empirical studies, and experimental validation, this thesis seeks to advance our understanding of the impact of program execution on neural program synthesis. By delving into the intricacies between program representation and behavior, we aim to pave the way for more robust and intelligent AI systems.