
Go game and loop integrals
At the ACAT workshop held in Tsukuba in 2003, Jos Vermasseren (NIKHEF, Amsterdam) compared the strategies in the game of Go and in solving equations for loop integrals in his presentation “The rules of physics“. Loop integrals are elements of computations appearing in computing higher order Feynman diagrams, at the kernel of any precise prediction for High Energy Physics (HEP).
Today an Artificial Intelligence (AI) program AlphaGo developed by Deepmind recently bought by Google proved to be able to beat one of the best professional Go player (Lee Sedol). This program actually combines Monte Carlo Tree Search (MCTS) and 12-layers convolutional neural networks. It uses an hybrid system based on conventional CPU’s and GPU’s. Would these methods and techniques described in the nature paper help in computing complicated loop integrals?
I was informed by Jos that he and his team have an ongoing HEP loop integrals research project, HEPGAME, using insights from AI like MCTS and Upper Confidence bounds applied to Trees (used in previous best Go game program). One of the main target here is to reduce the millions of terms originally produced which are intractable when it comes to the final numerical integrations. The use of deep neural networks, in AlphaGo, trained on millions of Go games, may proved to be challenging for loop integral calculations where it can be trained only on a limited pool of validated computations.


( J.Fleisher et al., https://inspirehep.net/record/1079778?ln=en)