Why an Optimized Plan Can Look Worse on One Measure

A proposal can show a worse number on one row because goals are ranked rather than blended: the optimizer will accept a small loss on a lower ranked measure to win a larger gain on the measure you asked for. That is not a flaw in the result and it is not hidden from you. The comparison table shows every measure with its current value, its proposed value, and a change cell that spells the direction out in words as well as an arrow.

EDGEBIC by User Solutions displays the trade rather than smoothing it away, because a planner deciding whether to disturb a floor needs to see the whole picture. This post explains how ranked goals produce these trades, how to judge one in a few seconds, and when a worse row means you picked the wrong goal.

Ranked, not weighted

Each goal is an ordered list of measures. On-time first ranks weighted lateness, then the count of late jobs, then stability, then plan span. Fastest overall finish ranks plan span first and lateness second. Fewest changes ranks stability first. Least setup time ranks changeover hours first.

Comparison then works by rank. Two plans are compared on the first measure where they meaningfully differ, and the winner on that measure wins outright. Lower ranked measures are only consulted when the higher ones tie. This is deliberately not a weighted score, because a weighted score with large multipliers can produce results nobody can explain: a big enough win on a minor measure eventually outvotes the thing you actually cared about. Ranking makes the boundary a real boundary. See how the optimizer ranks goals in strict order.

The consequence is exactly the behavior you are reading about. If a plan wins on your top measure, it is the winner even if it gives a little ground further down the list.

What the change column tells you

The comparison table has four columns: Measure, Current, Proposed, and Change. The change cell always states the direction in words alongside its arrow, such as a down arrow with "2 better", an up arrow with "0.5 worse", or a dash with "no change". Nothing depends on reading a color, which matters when the table is being read over a shoulder, printed, or projected in a meeting.

Read it in this order:

1. The verdict sentence above the table, which states in plain words what improved.
2. The top rows, which are the measures your chosen goal ranks highest.
3. Any worse row, and ask whether that measure matters more to you this week than the one that improved.

If the answer to step three is no, Accept. If it is yes, you picked the wrong goal, not a bad result.

Worked trades you will actually see

Late jobs improve, makespan gets slightly worse. You ran On-time first. The optimizer held a job back so an at-risk job could run earlier, and the plan tail extended a little as a result. Almost always a good trade in a date-driven plant.

Makespan improves, one job finishes later. You ran Fastest overall finish, which ranks plan span above lateness. It compressed the plan by running compatible work together, and one job with slack absorbed the shift. Good when you are freeing capacity, bad if that job was the tight one.

Setup hours improve, lateness slightly worse. You ran Least setup time. The optimizer grouped similar jobs to cut changeovers, which sometimes means holding a job back to run alongside its match. Good when changeover time is your real constraint. See how the optimizer cuts total setup time.

Everything improves at once. Common when the original sequence was starving a downstream machine. In the documented worked example, reordering three jobs took total lateness from three hours to zero, on-time jobs from one of three to three of three, and plan span from seventeen working hours to fourteen. Fixing starvation helps every measure, because idle time was the waste.

Why this is not a broken guarantee

The never-worse guarantee has a precise meaning: a proposal is never worse than your current plan on the goal you chose, judged by that goal's ranked measures in order. It has never claimed that every measure improves, because no real trade-off would allow that.

The guarantee is also enforced rather than promised. After the search finishes, the best candidate is compared directly against your current plan under the goal's ranking. If it does not beat the plan outright, your plan is what you are shown, along with a verdict saying so. A proposal that lost never reaches the Accept button. See the never-worse guarantee.

When a worse row means switch goals

A worse row is a prompt to check your goal against this week's reality. If your comparison table repeatedly shows makespan worsening and you are running a backlog clearance week, you are running the wrong goal. If lateness worsens week after week while setup hours improve, ask whether changeovers are genuinely your constraint.

Switching costs a run and a discard. Discard leaves the database exactly as it was, so you can run all four goals in a couple of minutes and compare what each is willing to trade. See comparing two optimizer runs before you commit.

One caution about setup hours

The setup hours measure is a flat total across the plan, summing the setup time recorded against each routing step. Once a sequence-dependent setup matrix is in play, the true setup cost of a plan depends on the realized sequence, so this column is an honest approximation rather than a sequence-aware total. Read it as a directional indicator, and lean on the Least setup time goal and the solver's native changeover handling for the real sequencing work.

The bottom line

A worse row means the optimizer made the trade your goal's ranking told it to make, and the table shows it plainly rather than burying it. Read the verdict, then the top rows for your goal, then decide whether the loss sits on something that matters more this week. If it does, discard and change goals rather than doubting the result. To see a comparison on your own data, open the Optimizer tab in Schedule Jobs and run. For more, read the EDGEBIC optimizer guide and how the optimizer scores a schedule, and explore the platform at EDGEBIC.