Manual experiments Automatic search Continuous demo MANUAL
Reveal surface New surface Clear experiments Next experiment ▶ Run ❚❚ Pause
0 — blue / minimum 50 100 — red / maximum
Click to conduct an experiment. The underlying surface is hidden until revealed.
Performance: best observed response vs. experimental budget Compare selected methods Clear comparison
Comparison methods run separately on the same surface and starting point with the same experiment limit. Their paths and progress curves are retained together. With noise enabled, each observation includes independent measurement error.
Search settings Optimization method Steepest ascent / descent Coordinate ascent / descent Quadratic RSM (CCD) Nelder–Mead simplex Simulated annealing Bayesian optimization (GP)
Objective Find maximum Find minimum
Number of local maxima 1 Generates separated, differently sized peaks.
Line-search step 5
Local design radius 5
Measurement noise, σ 0 Gaussian measurement noise; true surface remains unchanged.
Experiment budget 100
Animation speed 5
Exploration strength 1.5 Bayesian UCB exploration; annealing uses this to scale temperature.
Annealing cooling rate 0.96
Compare methods
Live results Experiments 0
Best observed —
Search stages 0
Last measured —
Ready for manual experiments.
Procedure log No experiments yet.
What is each algorithm doing? Steepest: fit a first-order plane through three local measurements; march along the estimated gradient until a step worsens the response; refit.
Coordinate: search along X and Y separately, reversing direction if necessary.
Quadratic RSM: collect a 9-run face-centered central composite design, fit six quadratic coefficients, and follow the predicted best response inside the local design region.
Nelder–Mead: move a three-point simplex via reflection, expansion, contraction, or shrinkage without a gradient.
Annealing: explore with stochastic moves, sometimes accepting worse measurements to escape local optima.
Bayesian: fit a small Gaussian-process surrogate with a fixed RBF kernel and use an upper-confidence-bound acquisition rule (or lower confidence bound for minimization).
Responses are kept as floating-point values for all calculations and only rounded for display. For noisy experiments, the measured response can fall outside the nominal 0–100 range; the true surface does not.