The Generate, Test, and Explain Discovery System Architecture

摘要

Discovery systems have the potential to make truly novel discoveries, although few of them actually do so. Because original discoveries are difficult to understand even when they are explained in lucid prose, discovery systems that make new discoveries should provide explanations of them that are easy for humans to understand, lest they go unnoticed. The generate, test, and explain discovery system architecture moves the machine discovery field a step closer to these two goals by: (1) Demonstrating on two real-world problems, financial time series analysis and secondary structure prediction, that discovery systems can deliver on their promise to make novel discoveries; and (2) Extending the autonomy of discovery systems with an explanation facility that rephrases discoveries in terms that are easy to understand.