A groundbreaking support system with reinforcement has been developed by a team of researchers from Google Deepmind. The system, named AlphaDev, has the unique ability to generate optimized algorithms without the need for prior training with examples of human code.
AlphaDev utilizes the same approach as Alphago, a program that is famous for learning to play games like GO, poker, and Starcraft. In this case, programming is viewed as a game where the objective is to minimize code delay while ensuring correctness.
The AlphaDev system was able to create new algorithms for essential computer science tasks, such as sorting and hashing data. For instance, the system succeeded in producing an algorithm for sorting the five elements that was 70% faster than the current human code. These algorithms were proudly included into C++, deeming them the first change in sorting algorithms for the language in over ten years. Faster hashing algorithms were also generated by the program, and were added to an open programming database.
The findings from the work of AlphaDev were published in the journal Nature. According to the authors of the article, the system may potentially help optimize computer code in conditions of physical restrictions of microchips. They also noted that the system currently only works with small programs and a subset of possible instructions. Thus, the question of scalability and applicability to more complex tasks remains unanswered.