Annealing or Perturbation Schedule¶
Procedure — instantiates Landscape-Aware Search Strategy Design
Allows controlled temporary worsening or variation injection to cross barriers, then gradually raises convergence pressure so the search settles into a good basin.
A search that only ever accepts improvements is a search that can never climb out of the first pit it falls into. The Annealing or Perturbation Schedule treats convergence pressure as a scheduled variable rather than a constant: early on the search is permitted to accept moves that make things temporarily worse — or to have deliberate variation injected — so it can cross ridges between basins; then, on a planned schedule, that permissiveness is tightened until the search cools into and refines whichever basin it has found. The one idea that makes it this mechanism and not a sibling is that all the diversity lives inside a single continuous trajectory whose acceptance rule changes over time — nothing is restarted, and no model is fitted.
Example¶
A university registrar is building next term's exam timetable: 900 exams into rooms and slots, minimizing student clashes and back-to-back sittings. A greedy scheduler that only ever swaps two exams when the swap immediately lowers the clash count jams almost at once — every remaining single swap makes things worse, even though a much better timetable exists two or three swaps away. The registrar's optimizer instead runs an annealing schedule. In the opening phase it accepts roughly a third of worsening swaps, letting the timetable wander uphill across the ridge that the greedy version could never cross. As the schedule cools over successive passes, the acceptance rate for worsening moves falls toward zero, and the search stops wandering and polishes. The illustrative result: the clash count that greedily plateaued around 140 settles near 30, and the registrar can point to the cooling curve as the reason it got there — the temporary worsening was budgeted, not accidental.
How it works¶
The procedure hangs on an acceptance rule governed by a falling control parameter (a "temperature"):
- Propose a move, evaluate the change in value.
- Accept unconditionally if better; accept probabilistically if worse, with the probability set by the current temperature — high temperature makes barrier-crossing easy, low temperature makes it nearly impossible.
- Cool on schedule. The temperature descends by a planned law (geometric, linear, or adaptive), shifting the run from exploration-dominant to exploitation-dominant.
- Optionally reheat if a trap indicator fires — a burst of perturbation amplitude to shake a stalled trajectory loose before cooling resumes.
The schedule is the mechanism: the same proposal engine produces wandering or convergence depending only on where the temperature sits.
Tuning parameters¶
- Initial temperature — how freely early worsening is accepted; too low quenches into the nearest pit, too high wastes budget on a random walk.
- Cooling rate — how fast permissiveness tightens; slower cooling reaches better basins but costs more evaluations.
- Perturbation amplitude — how large an injected variation is; large jumps cross wider barriers but blur local refinement.
- Reheat trigger — whether and when to inject a burst of temperature after a stall; adds trap-escape at the cost of settling time.
- Acceptance criterion — the exact worse-move acceptance law; sharper laws converge faster but forgive fewer detours.
When it helps, and when it misleads¶
It shines on rugged, path-dependent terrain where good solutions sit behind barriers that a strictly-improving search cannot cross, and where the cost of one more evaluation is low enough to afford some wandering. Its honest failure mode is the cooling schedule itself: cool too fast and you quench into a mediocre local optimum; cool too slow and you burn budget wandering long after the good basin was in reach.[1] The classic misuse is running a fixed textbook schedule regardless of terrain — treating the cooling law as sacred when the landscape's ruggedness should set it. The guarding discipline is to tie the schedule to evidence: watch how often worsening moves still lead somewhere, and let a stall indicator, not the calendar, decide when to reheat or stop.
How it implements the components¶
exploration_exploitation_balance— the temperature is the explicit explore/exploit dial, and the schedule makes its shift over time visible and deliberate.local_trap_indicator— a stalled trajectory that only escapes when perturbation amplitude rises is itself the trap signal this procedure reads and answers with reheating.strategy_switch_trigger— the scheduled temperature thresholds are pre-committed switch points from barrier-crossing to refinement.
It does not draw fresh starts across the configuration_space_scope — that relaunch-and-keep-best move belongs to Random Restart Plan, its nearest cousin; nor does it fit a surrogate_model (Response Surface Model) or lay out diverse landscape_probe_plan samples (Coarse Landscape Sampling).
Related¶
- Instantiates: Landscape-Aware Search Strategy Design — supplies the trap-escape-then-converge regime for rugged terrain.
- Sibling mechanisms: Coarse Landscape Sampling · Gradient or Directional Probe · Objective Surface Sketch · Optimization Trace Dashboard · Parameter Sweep and Sensitivity Grid · Random Restart Plan · Response Surface Model · Search Algorithm Portfolio
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Allows controlled temporary worsening or variation injection to cross barriers, then gradually raises convergence pressure so the search settles into a good basin, making its operative form a computation or analytic transformation that produces an inference, comparison, or optimized result.
Independent corroboration: The frozen evidence defines Annealing or Perturbation Schedule as 'Allows controlled temporary worsening or variation injection to cross barriers, then gradually raises convergence pressure so the search settles into a good basin', so its operative form is Analysis, Modeling & Optimization.
Nearest alternative: Experiment, Test & Rehearsal — The procedure computes a search result through a cooling optimization algorithm, while perturbation is internal to that computation.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Operations Research
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Scheduling temporary worsening followed by increasing convergence pressure is canonical simulated-annealing and metaheuristic optimization practice.
Related originating lineages:
- Chemistry & Materials Science — Physical annealing supplies the barrier-crossing and slow-cooling model.
- Computer Science & Software Engineering — Search algorithms implement perturbation and acceptance schedules.
- Mathematics — Markov-chain convergence provides formal analysis.
- Physics — Metallurgical cooling is the founding analogy.
Review resolution: The original simulated-annealing paper explicitly connects statistical mechanics to multivariate and combinatorial optimization, using a temperature schedule to search complex solution spaces. The page abstracts that optimization-control operation, making operations research primary while computing, probability, physics, and metallurgy remain materially formative.
Review outcome: Researched adjudication after independent review; high confidence.
Sources consulted:
References¶
[1] Kirkpatrick, S., Gelatt, C. D., Jr., and Vecchi, M. P. "Optimization by Simulated Annealing". Science 220(4598), 671–680 (1983). Explains that annealing quality depends on the cooling schedule and that rapid quenching can trap a search in a nonoptimal state. registry ↩