New AI Program discovers new solutions in maths and computer science: FunSearch
By: Alhussein Fawzi and Bernardino Romera Paredes
FunSearch is a new method that uses Large Language Models (LLMs) to search for new solutions in mathematics and computer science. Large Language Models (LLMs) are very good at combining concepts, and they can read, write and code to help people solve problems. The challenge has been that LLMs are not always accurate and may produce incorrect information. But FunSearch aims to use the LLMs’ creativity by identifying only their best ideas and building upon them.
The system works by pairing a pre-trained LLM with an automated evaluator to ensure accuracy and guard against incorrect ideas. FunSearch searches for “functions” written in computer code, hence the name FunSearch. This is the first time that LLMs have been used to make new discoveries for challenging open problems in science or mathematics.
The FunSearch process involves the LLM generating new programs, which are automatically evaluated. The best programs are then added back to the pool, creating a self-improving loop. Google’s PaLM 2 is used as the LLM for FunSearch, and the process is run in parallel to improve system efficiency.
FunSearch tackles challenging open problems in science and mathematics, providing new discoveries for problems like the cap set problem and the bin-packing problem. This system is breaking new ground in mathematics and may have practical applications in other fields. It also produces programs that describe how solutions were arrived at, making them easier for researchers to comprehend and providing actionable insights for further studies.