Two axes, two hyperparameters — say learning rate (x) and regularization strength (y). Underneath is a fixed, hidden validation-accuracy surface with several real local peaks: no strategy can see it directly, each can only query one (x, y) point at a time and learn that point's accuracy.
Grid search sweeps a fixed lattice of points, evenly spaced, in order — thorough but it spends equal effort on clearly bad regions and promising ones. Random search samples uniformly at random; a well-known finding (Bergstra & Bengio, 2012) is that for the same budget it often beats grid search, because it never wastes two trials on the same value of an unimportant parameter.
Bayesian-lite is a simplified surrogate-based optimizer: it starts by exploring broadly, then increasingly concentrates samples in a shrinking neighborhood around the best point found so far, with an occasional random "exploration" jump so it doesn't get stuck on a local peak — the same explore/exploit tension real Bayesian optimization resolves with a probabilistic surrogate model and an acquisition function.
The chart on the right tracks each strategy's best accuracy found so far against evaluations used — the curve that climbs fastest and highest is winning the trade-off for this landscape and budget.