ND Optimizer Arena

An interactive tool for understanding how popular optimization algorithms behave on complex loss surfaces. In 2D mode, you can compare multiple optimizers (SGD, Momentum, RMSProp, Adam, AdamW, Muon) simultaneously and observe their trajectories around local minima. For higher dimensions (3–15D), the tool displays a 2D slice of the loss landscape that updates dynamically as unobserved parameters change, letting you run one optimizer at a time.

Designed for students and practitioners learning optimization, the arena features procedurally generated terrain with adjustable hyperparameters like learning rate, batch size, and gradient clipping. The maker says it was built with Codex.