Catering production schedule — checked solution replay
TODAY’S PRODUCTION PLAN
PRODUCTION ORDER
Start order · separate cookers can overlap.
Prep → cook → pack is fixed. The plan chooses when, where and how batches overlap.
ORDERS · MAIN + SIDE
Discarded portions0
Click an order to jump to its departure.
Each program bit selects a batch quantity, cooker, crew and phase times. Each anchor block also has an idle bit. Exactly one bit in every block is 1.
Main and side choices interact through the complete-meal objective. Separate penalties forbid clashes between blocks. The XY mixer preserves one-hot choices; it does not eliminate every physical conflict.
The started Menu A batch is outside the register. We independently check every displayed continuation.
SCHEDULE BY BATCH
The kitchen, explained.
Orders, batch choices and constraints
In the small cell, one shared side batch completes two orders: the main already cooking for A, and a new main for B. Skip that side batch and both meals disappear from service, leaving two cooked mains to discard.
Choose batch sizes, shared allocations, cookers, workers and start times on a 10-minute grid. A meal needs one main course and one side course, both packed before its fixed departure.
Equipment and qualified workers cannot overlap. Freshness, calendars, prep, attended cooking, packing and cleaning constrain the plan. Started work stays committed. An outage also requires a worker to clean the repaired cooker before reuse.
One lot is 10 portions. U is priority-weighted missed complete-meal lots; W counts discarded component lots. The previous unstarted plan is empty in these fixtures, so R = B.
Ingredient stock, holding capacity, power, monetary cost and interrupted cooking are disabled in this version.
Paired-course objective
For meal d, Md and Sd are 0/1 coverage expressions for its main and side, including committed food.
One course alone produces waste. Pairing both serves the meal. Expanding these expressions gives quadratic coefficients with no extra product qubits.
Weights encode the declared priority order: each weight exceeds the maximum sum of lower-priority costs.
+ P ∑(i,j)∈conflictsxixj
Every anchor block contains batch programs and an idle choice. P = 1 + the sum of absolute nonconstant objective coefficients, which exceeds the objective range. The all-idle continuation is feasible; therefore a ground state has zero hard-constraint penalty.
QAOA applies cost phases and XY rotations within blocks. This preserves one-hot choices. Cross-block conflicts remain penalized and independently checked. Under x = (1 − Z)/2, the exported Ising coefficients represent exactly 4E.
Small complete cell. Larger restricted search.
Integrated method campaigns: all 400 QAOA runs are independently checked. The frozen configuration samples both held-out portfolio optima in 10/10 seeds at both budgets. Latest VQE uses CVaR→Gibbs concentration: median offered-optimum probabilities are 99.54%/99.55% on S2/S3, with 10/10 seeds each. These exact-gradient follow-up runs use previously examined cases; the older CVaR baseline remains available. Five new exports are independently checked. The staged gain did not carry over to shot-noise emulation. QAOA and VQE use unequal training budgets and final shots. Compare all campaigns, classical references and scaling ↗
Earlier retained experiments
Complete cell: 32 legal batch programs + 3 idle bits = 35 data qubits; 520 couplers. Its one-hot sector has 1,045 states, of which 59 are physically feasible. All six QAOA simulator trials find (0,0,2,2). Initializer and mixer-only controls also find that optimum.
Larger kitchen: four scenarios, two seeds, eight portfolio updates each, at most 35 qubits per circuit. Six of eight final plans attain the exact optimum of their final offered pool. All eight still trail the full-grid CP optimum, and final-step controls tie. This separates restriction of the choices from search within those choices.
Tuples are (U,W,B,R); smaller is better, compared left to right. These are exact-sector, exact-CVaR simulator experiments. They establish neither quantum advantage nor hardware readiness. Per-step controls share the quantum incumbent.
VeloxQ scale integration: proved idle elimination preserves feasible schedules and their objective while reducing 512/1,024/2,048 source bits to 486/998/2,022 solver bits. At equal 20,000-step budgets on the 512-bit source, the best checked plan serves 50 instead of 10 meals out of 140; discarded components rise from 30 to 60 portions because complete service has priority. Valid returns improve from 17/512 to 512/512; all three tested scales return 512/512 valid. These single-job representation measurements still trail CP and matched-formulation SA on quality. Inspect every service attempt, input identity and scope ↗
A saved result, replayed.
The QAOA view uses an actual retained sample, selected by checked objective and then frequency. The method selector changes the real schedule and handovers. Classical references and controls carry their own labels.
“Skip batch” is a checked manual intervention. Playback never calls a solver. Walking and art are illustrative; timing, worker attendance, meal handovers, discarded food and recovery cleaning come from the plan.