Google DeepMind’s AlphaZero AI, known for its mastery over board games like chess and Go, has now been utilised to generate sorting algorithms that outperform human-designed versions.
DeepMind’s AI system, called AlphaDev, has developed algorithms that can sort data up to three times faster when translated into the standard programming language C++.
Computer scientists have long worked to improve sorting algorithms in order to improve search results and organise data more efficiently.
DeepMind has achieved tremendous gains in sorting speed by applying the technology underpinning AlphaZero to sorting algorithms. These breakthroughs are detailed in a report published in Nature. AlphaDev first concentrated on sorting small integers spanning from 3 to 5 digits, which serve as building blocks for algorithms that sort larger lists. The AI system works at the assembly instruction level, improving code created by automatic compilers from programmer-written C++ code.
AlphaDev, like AlphaZero, blends reasoning with intuition. Rather than performing moves in a board game, it selects instructions to add to a method known as AssemblyGame. Neural networks, which are mathematical functions that learn from training data, assist AI in narrowing its search for the most promising commands.
AlphaDev is rewarded based on the accuracy and speed of its programmes. The system constantly modifies its neural networks based on game outcomes and experiments with other moves. In terms of sorting algorithms, AlphaDev pioneered two new strategies: the AlphaDev swap move and the AlphaDev copy move.
DeepMind researchers discovered considerable time savings in sorting and non-sorting algorithms. In terms of time efficiency, AlphaDev’s best sorting algorithms outscored human ones by 4 per cent to 71 per cent. AlphaDev achieved time savings of up to 67 per cent and 30 per cent for data conversion and hashing methods, respectively, for non-sorting algorithms.
While the scientific depth may not be surprising, experts say AlphaDev’s engineering prowess is impressive. DeepMind has proved its ability to adapt the AI system to effectively solve new issues, as demonstrated by the previous construction of AlphaTensor. Moreover, DeepMind envisions applying AlphaZero-style algorithms to a broader range of problems, including hardware design. The team aims to tackle the entire technology stack through further advancements in AI technology.