A team of researchers has trained an AI system called AlphaDev to discover and develop more efficient algorithms than those created by humans. The system optimises algorithms for measured latency at the CPU instruction level, using deep reinforcement learning (DRL) to search and more efficiently consider the correct programmes and their complexity. AlphaDev’s main application has been developing faster sorting algorithms; in the fixed sort setting, it has discovered two efficient sequences of instructions that, when applied to a sorting network algorithm, reduce the algorithm by one assembly instruction each time. The solutions discovered by AlphaDev include new algorithmic discoveries that lead to more efficient performance.
https://www.nature.com/articles/s41586-023-06004-9