Nelder–Mead Simplex Optimization

Optimization theory studies how to find the input that minimizes (or maximizes) a function, and not every real cost function is smooth enough to differentiate. This simulator renders a real 3D loss landscape — Himmelblau's function, the Rastrigin function, or a simple convex bowl — and drives a genuine Nelder–Mead downhill simplex search across it: a triangle of three trial points repeatedly reflects, expands, contracts and shrinks based only on comparing function values, exactly as the classic derivative-free algorithm defines it. Switch landscapes to see the same rule set converge quickly on a smooth bowl yet stall in a local dip on a rippled multi-modal surface, step through iterations one at a time, or let it run and watch the live iteration count, simplex size and best value converge.