Same exploration, different evaluation decisions

Synchronized LA paths · same random seeds · separate expensive f-call budgets
Ready
Round 0 / 60
Trial — / 50Waiting for first seed
Multistart-LA
Runs gradient-only LA, then always evaluates f at the endpoint.
BASELINE
Plot unavailable
Top-down: gradient paths / evaluated endpoints[-3,3]²
evaluatedbestcurrent seed/path
LipScreen-LA
Same LA endpoint, but evaluates only if it passes the Lipschitz screen.
SCREENED
Plot unavailable
Top-down: red excluded, green accepted endpointsAccepted: 100%
screen permitsscreen rejectsevaluatedbest
Default landscape: f(x,y) = Σ aⱼ exp(−((x−μⱼx)²+(y−μⱼy)²)/(2σⱼ²)) + 0.035 sin(1.7x+0.3) cos(1.5y−0.4). Other presets include 25 unequal Gaussian hills, rippled landscapes, and dense grids of local maxima. All smooth presets use analytic gradients; custom expressions use automatic differentiation. The plot uses the same 140 × 140 surface mesh as before. The objective plot is display-only: surface sampling is not charged against either evaluation budget.
What you should observe: both panels query the exact same gradients and finish each local ascent at the same endpoint. Multistart-LA spends one expensive function evaluation per run. LipScreen-LA instead updates k̂=max observed ‖∇f‖, takes ε=(1+γ)k̂, and skips an endpoint if minᵢ[f(yᵢ)+ε‖y−yᵢ‖] < M+ζ. The red/green map represents candidate endpoints, not seed rejection. Only gradient queries and accepted endpoint evaluations are charged to the algorithms; full objective plotting is reference-only. Some runs can decrease f outside locally concave regions, as permitted by the draft's assumptions.