Developer builds custom MLP evaluator for self-invented board game after rejecting NNUE

A software developer building an engine for Larss, an original 1v1 tetromino placement game, faced the challenge of creating an evaluation function with no existing theory or game data to draw from. Unlike chess engines, which benefit from decades of tuning and canonical implementations, Larss required the developer to bootstrap training data entirely through self-play. The developer considered using NNUE, the neural network architecture behind Stockfish's strength, but ruled it out due to Larss's variable board sizes (11×11 to 14×14) and the fact that tetromino placements can alter dozens of cells at once, undermining NNUE's incremental update efficiency. Instead, a multi-layer perceptron was designed to learn how to combine hand-crafted features such as score difference, mobility, and cell safety. Notably, the developer found that a simpler model ultimately outperformed the MLP, offering a practical lesson in matching model complexity to the problem at hand.