May 10, 2022
Keio University
A group of researchers, including Dan Kubota and Yudai Tokuoka (then graduate students at the Graduate School of Science and Technology, Keio University), Professor Kei Funahashi, and Senior Assistant Professor Takahiro Yamada of the Faculty of Science and Technology at Keio University, has developed an AI that can predict the integrated function (primitive function) from an input function to be integrated (integrand). They noted the similarity between the mathematical process of integration taught in high school and transformations such as language translation by AI, which has seen remarkable development in recent years. The group also devised a method to build and train various AIs and adopt the one that produced the correct answer, inspired by the idea that the correctness of an integration can be determined by differentiating the AI's output primitive function and checking if it matches the integrand. As a result, the implemented AI was shown to be capable of integration with an accuracy of 99.79%, achieving the highest precision compared to existing integration tools like Mathematica and other machine learning-based methods. Furthermore, by examining the characteristics of the mathematical formulas learned by the AI, it was revealed that each constructed AI has functions it is good or bad at integrating, and that the high accuracy was achieved by having the AIs solve integrations in a mutually supportive manner. Integration is an essential process for simulations in fields such as control engineering and systems biology, and this achievement is expected to contribute to more accurate simulations in these areas.
Ahead of its publication in the academic journal IEEE Access, an online preliminary version of this research was released on the journal's website on April 29.
Please see below for the full press release.