Benchmark
Parameters
the owning experiment's name
the problem case's name
the solver case's label
the 1-based macro-replication number
the unique cell label ("problem_solver_rN"); also the solver's name
the cell's lifecycle outcome (completed, failed, stopped before start)
the common starting point the cell's solver was given
the best design point the solver found (the problem's bad solution when the cell failed — see the isBestValid flag)
the estimated objective value at the best point (model scale)
the solver-internal penalized objective at the best point
false when the best solution is a placeholder from a failed or never-started cell
whether the best point satisfies the problem's input ranges and deterministic constraints
the total response-constraint violation at the best point (zero when response-feasible or the problem has no response constraints)
the best point's estimate for every response of the problem, keyed by response name, including the objective. These are what a selection rule needs in order to be replayed offline: with the average, variance and count of each response preserved, a solution sufficient for ranking can be rebuilt without re-simulating.
the best point's per-constraint assessment, in the problem's constraint order. The aggregate responseConstraintViolation above cannot say WHICH constraint bound; this can. Empty when the problem has no response constraints.
the solver's cumulative oracle-call count
the solver's cumulative requested replications — the actual budget consumption used for normalization
the solver's iteration count, when the solver ran
the solver's execution time in milliseconds, when tracked
the CPU time the cell's worker thread spent running the solver, in milliseconds. Portable across machines in a way wall clock is not, because it does not depend on the worker count or on what else the machine was doing. Counts only the worker thread, so work an evaluator delegates elsewhere is excluded. Null when the cell never ran or the JVM does not report per-thread CPU time.
the optimality gap of the best objective against the problem's gap basis
the basis the gap was computed against
the failure cause when the cell failed; null otherwise