A group of researchers from Google DeepMind in London found that artificial intelligence can find a faster algorithm to solve the matrix multiplication problem.
Mathematics is usually used in computer programming, usually as a means to describe and then manipulate the representation of real-world phenomena. For example, it is used to represent pixels on a computer screen, weather conditions or nodes on an artificial network.
In this case, one of the main methods to use mathematics is to use matrix to calculate. For example, when programming a game, the matrix describes the possible motion options.
In order to achieve this movement, matrices are usually multiplied or added-sometimes both actions are necessary. This requires a lot of work, especially as the matrix becomes larger and larger, so computer scientists spend a lot of time and energy to develop more and more efficient algorithms to complete the work.
For example, in 1969, the mathematician Volker Strassen discovered a method to multiply two 2×2 matrices using only seven multiplication operations instead of the standard eight multiplication operations.
In this new work, DeepMind researchers want to know whether it is possible to use artificial intelligence systems based on reinforcement learning to create new algorithms with fewer steps than those currently used.
In order to find the answer, they turned to the game system for inspiration and pointed out that most of them were based on reinforcement learning. After creating some initial systems, the team focused on tree search, which is also used for game programming. This is a way for the system to consider different schemes in a specific situation.
With regard to matrix multiplication, researchers have found that transforming an artificial intelligence system into a game allows you to find the most effective way to achieve the expected result-mathematical result.
Researchers have opened the search and browse functions to test their systems with existing algorithms, and used rewards as a temptation to choose the most effective algorithm to understand the factors that affect the efficiency of matrix multiplication.
Then, the researchers use the system to create their own algorithms and seek to improve efficiency again. The researchers found that in many cases, the algorithm selected by the system is better than the algorithm previously created by human beings.
In a paper published in the journal Nature, the team described the use of reinforcement learning to improve mathematical algorithms. In the same issue of the magazine, a research background document summarizing the group’s work in London was also published.