Inductive programming (IP) research – also called inductive program synthesis – is concerned with learning computer programs from data, typically input/output examples. It incorporates many areas of computer science, especially machine learning, automated reasoning, program verification, programming language theory, and software engineering. IP also has strong relations to research outside computer science, notably in cognitive science, where it can help build models of human inductive learning and intelligent tutoring systems for programming education. In industry, IP supports tools for end-user programming such as the Microsoft Excel plug-in FlashFill.

In contrast with most statistical and neural approaches to machine learning, IP approaches typically only need a small number of training examples and produce interpretable – also called explainable, symbolic, or white-box – models (i.e. programs). Because the learned programs can contain arbitrary computational structure (e.g. variables, loops, recursion), IP methods can be more expressive than other interpretable methods like linear regression and some forms of decision rules. IP also complements other approaches to program synthesis, including deductive and transformational techniques.

This Dagstuhl Seminar continues a series of previous seminars with the same title. The current seminar focuses on the relation between classic IP methods and generative AI approaches to code generation. Recent advances in generative AI use large language models (LLMs) to generate code from many types of specifications, including natural language descriptions, input/output examples, and partial programs. In contrast with classical IP methods, generative approaches are highly flexible: they can often accept many types of specifications, generate code in multiple programming languages, and be used for multiple related tasks (e.g. code generation, example generation, code repair, and translation between programming languages and natural language). On the other hand, generated code is often incorrect and of low quality. This seminar aims to explore ways to combine traditional IP methods with new techniques from generative AI for increased performance, flexibility, and robustness. We also plan to address efficient algorithms for IP, especially inductive logic programming (ILP); methods for evaluating the quality of induced code; techniques for integrating IP with cognitive models of learning and programming education; and IP’s usefulness in domains that combine syntactic pattern matching and semantic reasoning.

One long-term objective of the Dagstuhl Seminar series is to establish IP as a self-contained research topic relevant to AI, ML, and cognitive science. The seminar thus serves as a community-building event by bringing together researchers from different areas of IP, mainly inductive functional programming, inductive logic programming, rule learning, and generative AI for code generation. Furthermore, it also regularly includes insights from industry researchers interested in building IP-based tools for end-user programming and human-AI co-creation of code. Researchers from other areas of AI, interested in learning expressive rules, are invited, such as learning in planning and learning and theorem proving, and learning in the context of knowledge graphs. In addition, researchers from the areas of (transformational and deductive) program synthesis, and from logic and functional programming are included to explore combinations with IP methods. Finally, researchers from cognitive science, psychology, and education help these seminars to explore relations between human learning and IP, including applications in the domain of programming education.

Creative Commons BY 4.0

Sebastijan Dumancic, Josh Rule, Ute Schmid, and Gust Verbruggen