Engineering methods
Reproducible Numerical Optimization for Engineering
Declare the objective, budget and acceptance checks before comparing solvers. A bounded analytic example separates an attractive number from a feasible result.
Write the acceptance question first
An optimizer answers the numerical problem it receives. It does not establish that the objective matches an operational need, that units are consistent or that an actuator can safely execute the answer. Start by writing the decision variables, objective, bounds, constraints and the meaning of acceptance.
Optimization Engineering makes this separation explicit. Its public reference experiments and archived publication document a workflow; they are not a blanket claim of industrial certification or universal solver superiority.
Use a problem with an independent answer
A small analytic problem can test the validation boundary before a complicated model hides mistakes. Minimize f(x) = (x - 2)^2 subject to -1 <= x <= 1. The unconstrained minimum is x = 2, but that point is infeasible. On the feasible interval the minimum is x = 1, where the objective is 1.
def objective(x):
return (x - 2.0) ** 2
def verify(x):
feasible = -1.0 <= x <= 1.0
return {'x': x, 'feasible': feasible, 'objective': objective(x)}
assert objective(1.0) == 1.0
assert verify(2.0)['objective'] == 0.0
assert not verify(2.0)['feasible']
assert verify(1.0)['feasible']The point with the smallest printed objective can therefore be the wrong engineering answer. A solver success message and an independent feasibility decision should be stored as separate fields.
Declare a fair comparison budget
Wall-clock time, objective evaluations, gradient evaluations and memory measure different resources. A method using an analytic gradient and one using only objective values do not receive identical information simply because their iteration counts match.
Declare the comparison resource and record the others. Keep the objective implementation, initial conditions, tolerances and termination criteria in the experiment record. A trace should preserve the best observed feasible value at each declared checkpoint. Extending a terminated curve for display should be labeled as a visual convention, not a new measurement.
Separate tuning from assessment
If you choose a configuration using the same random seeds later used to report success, the assessment is influenced by selection. Keep development seeds and held-out seeds separate. State the baseline and improvement threshold before the final comparison. For deterministic problems, vary the conditions that support the intended claim rather than manufacturing random repetitions of an identical calculation.
REI Scientist provides an offline-first implementation of bounded synthetic studies with declared protocols and independent verification. Its reference results concern specified benchmark runs. They do not establish performance on an industrial dataset or constitute independent replication.
Preserve enough evidence to replay
Save the study definition, runtime versions, bounds, seeds, objective accounting, candidate result and verification outcome. Preserve failures and budget exhaustion. A configuration that fails to converge is still informative evidence about the declared experiment.
Integrity hashes can detect changes to saved artifacts, but unsigned hashes do not authenticate the original publisher. Re-executing a recorded run checks a different property from checking file integrity; record both results explicitly. If numerical differences occur, investigate tolerances and the runtime environment before declaring them either harmless or a regression.
Acceptance checklist and public references
- Is the objective connected to an explicit engineering requirement?
- Are bounds, units and constraints independently checked?
- Was the resource budget declared before comparison?
- Is configuration selection separated from held-out assessment?
- Are failed and terminated runs retained in the evidence?
- Does the claimed conclusion stay within the tested problem set?
Inspect the Optimization Engineering validation protocol, the REI Scientist reference studies and the versioned Zenodo publication. A numerical result can support a design decision while still requiring an independent system-specific acceptance process.
Explore Rootcastle engineering tools → · Public profiles and publications
