Scientists from the artificial intelligence (AI) lab at Tinkoff Research have discovered a new algorithm for AI training.
The new method, called SAC-RND, trains robots 20 times faster and 10% better than any existing analogue. Such results have been demonstrated by testing the method on robotic simulators.
The new algorithm can improve the safety of unmanned vehicles, simplify supply chains, speed up the delivery and operation of warehouses, optimize combustion processes in power plants and reduce emissions of harmful substances to the environment, as well as bring earth science closer to creating a new solution. Universal robot that can perform any task alone.
The results of the research were recognized by the global scientific community and were presented at the 40th International Machine Learning Conference (ICML), along with the latest achievements from Google DeepMind, Amazon, Sony and other tech giants.
It was previously believed that the use of random neural networks (sequential and automated decision-making algorithms, RND) is unsuitable for offline training of reinforced robots. Tinkoff Research refuted the results of previous experiments and managed to optimize the method, making it the most effective among all available analogues.
Scientists from Tinkoff Research are exploring the most promising areas of artificial intelligence: natural language processing (NLP), computer vision (CV), reinforcement learning (RL), and recommendation systems (RecSys). Scientific work from Tinkoff Research is cited by scientists from Berkeley and Stanford universities, as well as by the Google DeepMind artificial intelligence research project.