The developer created an AI-powered version of the classic paper-and-pencil game Dots, now available as a free demo on Steam under the title Just Dots. The AI was trained using an approach inspired by AlphaZero, where the neural network learns by playing games against itself.
The training process involved 30 generations of the model over one and a half months. Training time varies significantly based on board size, ranging from approximately 3 hours for 9x9 and 11x11 grids to 35 hours for 25x25 grids. The model is built using PyTorch and exported to the ONNX format for integration into the C# game engine.
The game features 15x15 and 20x20 boards with three difficulty levels. While the AI uses a neural network for complex play, the easy mode relies on a separate heuristic-based algorithm. The developer plans to add larger boards, including the classic 39x32 size, and a fourth difficulty level, with a full release scheduled for December.
Technical challenges during development included optimizing the search budget and tree caching. While caching improved performance, it initially hindered training by increasing the frequency of draws, which reduced the number of winning examples available for the model to learn from. Consequently, caching is disabled during the training phase.
