The company DeepMind, known for its advancements in artificial intelligence and neural networks capable of playing computer and board games at a human level, has unveiled the AlphaCode project, which develops a machine learning system for code generation that can participate in programming competitions on the Codeforces platform, achieving average results. A key feature of this development is its ability to generate code in Python or C++, taking in text problem statements in English.
For testing the system, 10 new Codeforces competitions with over 5000 participants were selected, conducted after the machine learning model training concluded. The results of the tasks allowed AlphaCode to rank approximately in the middle of the specified competitions (54.3%). The predicted overall rating for AlphaCode was 1238 points, placing it in the Top 28% among all Codeforces participants who have competed at least once in the last 6 months. It is noted that the project is still in its early stages of development, and there are plans to improve the quality of the generated code, as well as evolve AlphaCode into systems that assist in writing code or application development tools usable by individuals without programming skills.
The project utilizes the 'Transformer' neural network architecture combined with sampling and filtering methods, enabling the generation of various unpredictable code variations that correspond to natural language text. After filtering, clustering, and ranking, the most optimal working code is selected from the generated stream of options, which is then verified for producing the correct result (each competition task specifies a set of input data and the corresponding output that should be achieved after running the program).
For the preliminary training of the machine learning system, a codebase available in public GitHub repositories was used. After preparing the initial model, an optimization phase was carried out based on a collection of codes with examples of tasks and solutions proposed to participants of competitions such as Codeforces, CodeChef, HackerEarth, AtCoder, and Aizu. A total of 715 GB of code from GitHub and over a million example solutions to typical competition tasks were utilized for training. Before moving on to code generation, the problem statement underwent a normalization phase, during which all extraneous elements were removed, leaving only the significant parts.
Source: opennet.ru
