The 'FunSearch” training method announced by Google DeepMind: enables AI models to solve complex discrete mathematical problems

According to news on December 15, Google DeepMind recently announced a model training method called "FunSearch", which claims to be able to calculate "upper-limit problems" and "boxing problems" A series of "complex problems involving mathematics and computer science."

The content that needs to be rewritten is: ▲ Source: Google DeepMind (hereinafter referred to as DeepMind)

It is reported that the FunSearch model training method is mainly introduced for AI models An "Evaluator" system is developed. The AI ​​model outputs a series of "creative problem-solving methods", and the "Evaluator" is responsible for judging the problem-solving methods output by the model. After repeated iterations, it can be trained Develop AI models with stronger mathematical capabilities.

Google's DeepMind used the PaLM 2 model for testing and established a dedicated "code pool" to input a series of questions in the form of code and set up the evaluator process. The model then automatically selects problems from the code pool, generates "creative new solutions" in each iteration, and submits them to the evaluator for evaluation. Among them, the "best solution" will be re-added to the code pool and start another round of iteration

This site noticed that the

FunSearch training method is particularly good at "Discrete Mathematics (Combinatorics)". After training The model trained by this method can easily solve extreme value combinatorial mathematics problems. In a press release, the researchers introduced the process of calculating the "upper-level problem (a central problem in mathematics involving counting and permutations)" by the model. .

Moreover, the research team also successfully solved the "Bin Packing Problem" using FunSearch training technology. This problem refers to how to fit items of different sizes in the smallest number of containers. FunSearch provides a real-time solution and generates a program that automatically adjusts based on the actual volume of the item

The researchers mentioned that

is different from other exploits Compared with the AI ​​training method of neural network learning, the output code of the model trained by the FunSearch training method is easier to check and deploy, which means it is easier to be integrated into the actual industrial environment.

The above is the detailed content of The 'FunSearch” training method announced by Google DeepMind: enables AI models to solve complex discrete mathematical problems. For more information, please follow other related articles on the PHP Chinese website!