The TinyBrains contest has been launched to optimize neural networks in strategic games

Organizers presented TinyBrains — a competition for creating compact neural networks capable of efficiently managing game units in the Ants strategy. Participants are invited to develop algorithms that demonstrate high performance under strict model size constraints.

A key condition of the competition is compliance with weight categories, where the maximum size of neural network files is limited to 64 MB. The organizers have implemented a leaderboard system that ranks participants not only by game results but also by the efficiency indicator per byte of model weight. This encourages developers to use quantization, pruning, and knowledge distillation methods.

The Ants game requires the neural network to make decisions in real-time, making the task challenging for models with a limited number of parameters. Participants must ensure the stable operation of the agent in the dynamic game environment, where each action requires minimal computational costs. The source code and participation rules are available on the project's official website.

The competition is aimed at developing neural network compression methods for running on devices with limited resources. Details of the discussion and technical implementation details can be found in the thread on Hacker News. The project emphasizes the importance of finding a balance between decision quality and architecture compactness in reinforcement learning tasks.