Directional subset simulation method for reliability analysis

Abstract: Estimating the probabilities of rare failure events is a key challenge in the reliability analysis of physical systems. Subset simulation (SS) is a very popular adaptive Monte Carlo method for this problem. In SS, the small failure probability is evaluated as a product of larger conditional probabilities by iteratively sampling a sequence of nested sub-domains of the parameter space, encompassing the target failure domain of interest, using Markov chain Monte Carlo methods. For failure domains with multiple modes, the Markov chain samples used to explore the intermediate levels of SS can be trapped in a confined region of the input parameter space, leading to inaccurate failure probability estimates. In this contribution, we propose the directional subset simulation (dSS) method for this problem, which uses concepts from directional sampling to informedly propagate samples towards failure. This is accomplished through a novel selection of the intermediate failure domains, which preserves samples in several directions in the parameter space in each intermediate level. The merits of the dSS method are illustrated through a selection of numerical examples.

• The paper introduces directional subset simulation, which uses adaptive, bin-specific thresholds and residual resampling to preserve exploration across multiple failure modes.
• The method substantially improves reliability estimation on multimodal benchmarks, reducing log-scale RMSE from 1.89 to 0.50 at N=500 while producing per-mode probability estimates.
• The results show that dSS depends strongly on informed directional partitions, with higher-dimensional binning, increased variance, and limited two-dimensional validation remaining important constraints.

Motivation and problem setting

The paper addresses the estimation of small failure probabilities $P_F = P(g(\Theta) \le 0)$ in reliability analysis, where the limit-state function (LSF) $g$ defines a failure domain $F$ in an $n$-dimensional parameter space. The central difficulty considered is that of multi-modal failure domains: when the truncated density $$p(\theta \mid F) \propto p(\theta)\mathbf{1}_F(\theta)$$ has several dominant regions separated by low-probability areas, standard subset simulation (SS) can fail. In SS, Markov chain Monte Carlo (MCMC) samples at intermediate levels may become trapped in a confined region of the parameter space—typically because of a steep local gradient in the LSF—so that entire failure modes go undetected, producing severe underestimation of $P_F$ or high sampling variance.

Brute-force Monte Carlo is unbiased but requires sample sizes scaling as $1/P_F$; FORM/SORM handle multiple modes only through extensions around several design points; and existing remedies to SS's mode-trapping problem have been proposed but do not resolve it structurally. The authors' response is a new variant, directional subset simulation (dSS), which borrows the idea of exploring directions from directional simulation while retaining the multi-level conditional decomposition of SS.

The dSS method

dSS generalizes SS by replacing the single global threshold at each level with bin-wise constant thresholds. The parameter space $\mathbb{S}$ is partitioned into $J$ bins $(B_j)_{1\le j\le J}$, each a linear cone of half-lines emanating from the origin (hence "directional"). Within each bin, a nested sequence of local subsets $g$0 is maintained, and the global intermediate failure domain at level $g$1 is

$g$2

Standard SS is recovered as the special case $g$3. Local thresholds are selected adaptively: $g$4 is set to the empirical quantile of order $g$5 of the LSF values of the particles currently in bin $g$6, provided this estimate is positive; otherwise the bin has reached the limit state and is retired from further exploration. Because each bin retains its own threshold, samples are preserved in several directions simultaneously, preventing the collapse onto a single mode that afflicts SS.

Several algorithmic details distinguish dSS from a naive stratified combination:

• Residual resampling: since the number of seeds $g$7 falling in the active region is close to, but not exactly, $g$8, residual resampling determines how many MCMC offspring each seed produces, reducing to classical SS behavior when $g$9 exactly.
• Simultaneous bin exploration: all bins are advanced concurrently rather than sequentially, which enables a principled stopping rule.
• Per-bin probability estimates: for a retired bin $F$0, the failure probability contribution is estimated as $F$1, where $F$2 is assumed known (equal to $F$3 in the experiments). For active bins, an upper bound $F$4 is available, and the algorithm terminates when the total unretained upper bound falls below $F$5 times the accumulated estimate.

An important structural assumption is that the bin probabilities $F$6 are known analytically—a condition satisfied by cones under the standard Gaussian measure after iso-probabilistic transformation, but one that constrains admissible partitions.

Numerical evidence

Two two-dimensional benchmark problems with known reference values from $F$7-sample brute-force MCS are studied, using $F$8 independent runs per configuration.

Piecewise-linear series system ($F$9)

This LSF exhibits a dominant mode and a secondary mode with a steeply decreasing gradient near the origin. Standard SS loses the dominant mode from the second level onward and grossly underestimates $n$0 in most runs. With a partition aligned with the failure structure (Cases 1 and 2), dSS keeps both modes populated at every level. Representative results:

Method	$n$1	Mean estimate	CoV	RMSE $n$2
SS	500	$n$3	2.90	1.89
dSS (Case 1)	500	$n$4	1.33	0.50
SS	4000	$n$5	0.92	0.71
dSS (Case 1)	4000	$n$6	0.40	0.17

At $n$7, dSS reduces the log-scale RMSE by roughly a factor of four relative to SS at comparable LSF evaluation counts ($n$8 vs. $n$9). The estimates carry a slight positive bias that diminishes as $$p(\theta \mid F) \propto p(\theta)\mathbf{1}_F(\theta)$$0 grows. Critically, performance depends on the partition: in Case 3, where the dominant mode shares a bin with half of the secondary mode, dSS performs no better than SS.

Four symmetric beta points ($$p(\theta \mid F) \propto p(\theta)\mathbf{1}_F(\theta)$$1)

For $$p(\theta \mid F) \propto p(\theta)\mathbf{1}_F(\theta)$$2, there are four equally important failure modes. Here SS does not underestimate $$p(\theta \mid F) \propto p(\theta)\mathbf{1}_F(\theta)$$3 (equal mode weights compensate), but it fails to identify all modes. dSS explores all four (or eight, with finer partitioning) modes, at the cost of somewhat higher variance and RMSE—for instance, at $$p(\theta \mid F) \propto p(\theta)\mathbf{1}_F(\theta)$$4, dSS Case 1 yields CoV 1.45 vs. 1.02 for SS—because the same per-level budget must estimate multiple local thresholds. The authors present this explicitly as a trade-off: reduced global-estimation efficiency in exchange for complete mode identification and per-mode probability contributions.

Limitations and open questions

The paper is candid about three constraints. First, performance is partition-dependent: a poorly chosen binning (Case 3 of Example 4.1) eliminates the advantage over SS entirely, so expert knowledge of failure modes is effectively required for reliable gains. Second, the curse of dimensionality in binning: the default orthant partition produces $$p(\theta \mid F) \propto p(\theta)\mathbf{1}_F(\theta)$$5 bins, each requiring its own threshold estimate from a shared sample budget, which degrades variance—the eight-bin Case 2 of Example 4.2 already shows elevated CoV. Third, the numerical validation is limited to two-dimensional problems, so behavior in higher dimensions remains untested. The proposed remedy—an adaptive mechanism that detects multimodality via clustering on LSF values and splits bins dynamically—is outlined only as future work, as is a systematic comparison against other SS variants.

Conclusion

dSS extends subset simulation by allowing bin-wise constant thresholds within directional cones, ensuring that all designated directions remain populated across intermediate levels. On benchmarks where SS's Markov chains collapse onto a single mode, dSS restores full exploration of the failure domain and substantially reduces log-scale RMSE (e.g., $$p(\theta \mid F) \propto p(\theta)\mathbf{1}_F(\theta)$$6 vs. $$p(\theta \mid F) \propto p(\theta)\mathbf{1}_F(\theta)$$7 at $$p(\theta \mid F) \propto p(\theta)\mathbf{1}_F(\theta)$$8), while providing per-mode failure probability estimates. Its effectiveness hinges on informed bin selection, and its extension to adaptive binning and higher-dimensional problems constitutes the principal open question left by the paper.