Find your next learning step
Several queue proposals can cover different rare-delay mechanisms, but drawing from many proposals is only half the job. Multiple importance sampling also specifies how every path is weighted against the proposal collection.
Open the chapter contents · Queueing Theory root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the question, or read from Chapter 1 for the complete argument. The chapters remain open on the page; these grouped routes are an index.
Route 1 · Frame the problem
Route 2 · Work the ledgers
Route 3 · Transfer the method
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · FRAME THE PROBLEM
1. Why proposal diversity and weighting are separate decisions
A queue simulation may use one proposal for bursty arrivals, another for slow service and a third for recovery failures. Those proposals determine which paths are generated. The multiple-importance-sampling, or MIS, rule determines how their evidence is combined.
A valid mixture estimator evaluates the target density against a declared denominator. It does not let each path choose whichever denominator gives the largest contribution. Proposal selection and weight construction must be fixed before examining the answer, or handled by a formally adaptive method.
CHAPTER 2 OF 12 · FRAME THE PROBLEM
2. One-sample mixtures and deterministic allocation
In a one-sample mixture, a component label is drawn using shares αj and then a path is drawn from that component. The marginal sampling density is qmix=Σαjqj. A standard return weight uses the baseline density divided by qmix.
In deterministic multiple sampling, the analyst fixes nj draws from each proposal. Balance-heuristic weights use the same proposal counts inside the denominator. This can reduce random variation in component counts, but the estimator must match the allocation actually used.
CHAPTER 3 OF 12 · WORK THE LEDGERS
3. A proposal-and-threshold ledger
Suppose 12,000 paths are allocated across three specialised proposals. Each proposal reaches a different part of the delay curve. The table records outcomes before any final weighting.
| Proposal | Paths | 30-minute hits | 60-minute hits |
|---|---|---|---|
| Arrival burst | 5,000 | 620 | 42 |
| Service slowdown | 4,000 | 390 | 68 |
| Recovery failure | 3,000 | 170 | 91 |
| Total | 12,000 | 1,180 | 201 |
The recovery proposal supplies fewer moderate-delay hits but nearly half of the 60-minute hits. Pooling raw hit rates would therefore overstate the baseline tail and discard the density information that makes importance sampling valid.
CHAPTER 4 OF 12 · WORK THE LEDGERS
4. A balance-heuristic contribution ledger
For one 60-minute path, suppose the baseline density is 0.002 and the proposal densities are 0.010, 0.004 and 0.020. With sampling shares 5/12, 4/12 and 3/12, the combined proposal density is about 0.01083, giving a baseline-return ratio near 0.185.
| Denominator choice | Value | Return ratio | Interpretation | |
|---|---|---|---|---|
| Generating proposal only: 0.020 | 0.020 | 0.100 | Ignores coverage by other proposals | |
| Equal proposal average | 0.01133 | 0.01133 | 0.176 | Wrong if sample shares are unequal |
| Count-weighted mixture | 0.01083 | 0.01083 | 0.185 | Matches the declared allocation |
| Smallest proposal density | 0.004 | 0.004 | 0.500 | Data-driven choice inflates influence |
Balance weighting recognises that a path well covered by several proposals should not receive separate full-strength contributions from each. The result is robust combination, not permission to change proposals after seeing the path.
CHAPTER 5 OF 12 · WORK THE LEDGERS
5. Deterministic-mixture and proposal-specific denominators
Several unbiased or consistent MIS constructions exist under their stated sampling models. Some use the full deterministic mixture for every draw; others use proposal-specific terms with explicit combination weights. Their variances and computation costs differ.
For queue paths, evaluating every proposal density can be expensive because the likelihood factorises across many arrivals and services. Store stable log densities and use log-sum-exp arithmetic. A cheaper denominator is not an improvement if it no longer represents the design.
CHAPTER 6 OF 12 · TRANSFER THE METHOD
6. Threshold-spanning evidence is a vector, not one score
For each delay threshold, decompose the final estimate by proposal, inspect the largest weighted contributions and calculate replicate uncertainty. A proposal can dominate because it is uniquely useful or because another density was evaluated incorrectly. The ledger alone does not decide which explanation is true.
Use a threshold ladder that shows probability estimate, component contributions, effective count, Pareto-tail diagnostic and action margin. Stop extrapolating at the first threshold where the estimate is carried by a handful of unstable paths.
CHAPTER 7 OF 12 · TRANSFER THE METHOD
7. Changed-context transfer: network incident queues
Transfer the design to a network-service queue. One proposal creates request bursts, another lengthens service times and a third increases retry feedback. MIS can combine these routes to a latency breach while retaining the baseline target.
The transfer requires a joint path likelihood for every altered mechanism. If retry behaviour depends on congestion, separate tilts cannot be multiplied as though independent. The operational loss threshold and recovery policy also need domain-specific modelling.
CHAPTER 8 OF 12 · TRANSFER THE METHOD
8. Connecting MIS to the preceding queue guides
The mixture-proposal guide designs support across mechanisms and thresholds. This article narrows in on the estimator that combines those proposals. The Pareto-tail guide then audits the most influential return ratios.
Use likelihood-ratio diagnostics and tilt stability to check direction, concentration and replicate stability before interpreting any capacity recommendation.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and misconceptions
The balance heuristic is not automatically variance-optimal for every finite design. Better weights may exist, but estimating them can add instability and optimisation error. A lower Monte Carlo variance cannot repair a wrong baseline model, missing support or a reversed likelihood ratio.
More proposals can waste runs when their coverage overlaps heavily. A proposal that produces many tail hits may still contribute little after valid weighting. Raw hit counts, component shares and final weighted influence answer different questions.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Practical evidence checklist
Archive the baseline density, every proposal density, sample counts, allocation rule, component labels, random seeds and per-path log densities. Verify normalisation, support and denominator construction on analytically tractable cases.
Report threshold-specific component contributions, raw and normalised weights, effective counts, largest shares, tail diagnostics, replicate error and action switching points. Recompute with a simpler valid estimator and with selected proposals removed to test whether the conclusion depends on one component.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Frequently asked questions
Is the balance heuristic the same as averaging proposal estimates?
No. It weights path contributions using the proposal collection and sampling counts. Averaging separately normalised estimates targets a different construction.
Must every density be evaluated for every path?
Full-mixture balance weighting normally requires it. Special structure or another valid MIS scheme may reduce work, but the estimator must be derived explicitly.
Can the best proposal be chosen after the run?
It can be reported as an exploratory diagnostic. Reusing the same outcomes for unaccounted selection and final inference can bias the result.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Sources and next reading
Veach and Guibas introduce multiple importance sampling and the balance heuristic. Elvira and colleagues compare efficient MIS estimators. Owen and Zhou develop safe and effective importance sampling. The queue ledgers above are original teaching examples.
Continue through the Queueing Theory root, mixture proposals and defensive sampling and Pareto-tail diagnostics. The Gold Standard of Education and How X Works provide the wider route.
Top · Contents · Previous · How X Works