The bumpy surface is a synthetic hyperparameter loss L(λ) over two axes (log learning-rate, log regularization). A marker walks the surface using the selected strategy while every visited point is left behind as a trail. Regret_T = Σ_t=1^T [L(λ_t) − L(λ*)] Random Search samples λ_t uniformly. Online GD follows −η·∇L(λ_t), perturbed by adversarial noise δ (robust optimization: min_λ max_δ L(λ,δ)). Meta Warm-Start begins from a pre-learned good region — λ*_new = MetaLearn({(D_i, λ*_i)}) — then fine-tunes, so its regret grows far slower than the other two.