Additive OCBA for Optimal Budget Allocation

• Additive OCBA is a framework for optimal computing budget allocation that decomposes the overall selection error into an additive sum of critical pairwise comparisons.
• It reduces complex robust ranking-and-selection problems by focusing simulation efforts on k + m - 1 effective scenarios, thereby simplifying the allocation strategy.
• Sequential additive OCBA algorithms incrementally allocate simulation resources based on variance-to-mean gap ratios to achieve asymptotically optimal performance.

Additive OCBA denotes a family of optimal computing budget allocation formulations in which the budget-allocation objective is expressed through an additive decomposition of selection error, or the allocation rule is implemented additively through sequential sample increments. In the most explicit usage, robust ranking and selection under input uncertainty is reformulated by an additive upper bound on the probability of incorrect selection, reducing a $km$-scenario robust problem to a classical OCBA problem with $k+m-1$ effective alternatives (Wan et al., 2024). Closely related work studies additive or sequential OCBA rules that repeatedly add samples to designs according to OCBA optimality conditions (Li et al., 2022), and budget structures in which simulation time rather than replication count is the primitive resource, with estimator variance behaving as $\sigma_i^2 \mu_i / T_i$ asymptotically (Jia, 2012). Taken together, these results position additive OCBA as an extension of classical OCBA to structured error decompositions and structured budget accumulation.

1. Conceptual position within the OCBA literature

Optimal computing budget allocation (OCBA) is a fixed-budget ranking-and-selection methodology that allocates simulation effort across a finite set of designs to maximize the probability of correct selection or, equivalently in large-sample formulations, to minimize the probability of false or incorrect selection. In the classical setting analyzed in convergence-rate form, there are $k$ designs, design $i$ generates i.i.d. samples $X_{i,l} \sim \mathcal{N}(\mu_i,\sigma_i^2)$, the true best design is $b = \arg\max_i \mu_i$, and the total sampling budget is $n$ with proportions $\alpha_i = n_i/n$ (Li et al., 2022).

Within this literature, “additive OCBA” appears in two technically distinct but related senses. First, in robust ranking and selection, the probability of incorrect selection can be upper-bounded by an additive sum of pairwise comparison probabilities involving only a small critical subset of scenarios, yielding a reduced-dimensional OCBA problem and the AR-OCBA procedure (Wan et al., 2024). Second, sequential OCBA algorithms such as OCBA-1 and OCBA-2 are additive in the operational sense that they repeatedly add one batch of samples to a selected design so that empirical allocations track the OCBA proportions (Li et al., 2022).

A plausible implication is that additive OCBA is best understood not as a single algorithmic identity, but as an OCBA design principle: express the dominant selection-error mechanism in additive form, then allocate increments of budget to the components that appear in that additive representation. The stochastic-time extension OCBAS reinforces this view by showing that when the primitive budget is total simulation time, the variance term $\sigma_i^2/n_i$ is replaced asymptotically by $k+m-1$0, so classical OCBA structure can be retained after an effective-budget transformation (Jia, 2012).

2. Robust ranking and selection formulation

The direct formulation of additive OCBA arises in robust ranking and selection under input uncertainty. The setting contains a finite set of alternatives

$k+m-1$1

and a finite ambiguity set of candidate input distributions

$k+m-1$2

For alternative $k+m-1$3 and distribution $k+m-1$4, the mean performance is $k+m-1$5, and the robust objective is

$k+m-1$6

The paper represents each alternative-distribution pair as a scenario $k+m-1$7, with scenario outputs $k+m-1$8, mean $k+m-1$9, and variance $\sigma_i^2 \mu_i / T_i$0 (Wan et al., 2024).

The analysis adopts the mean-configuration assumption

$\sigma_i^2 \mu_i / T_i$1

for each alternative $\sigma_i^2 \mu_i / T_i$2, together with

$\sigma_i^2 \mu_i / T_i$3

Under this assumption, scenario $\sigma_i^2 \mu_i / T_i$4 is the worst-case scenario for alternative $\sigma_i^2 \mu_i / T_i$5, and alternative $\sigma_i^2 \mu_i / T_i$6 is the unique robust best alternative (Wan et al., 2024).

With total simulation budget $\sigma_i^2 \mu_i / T_i$7, let $\sigma_i^2 \mu_i / T_i$8 be the final sample size for scenario $\sigma_i^2 \mu_i / T_i$9, and let $k$0 be the sample mean. The estimated worst-case mean of alternative $k$1 is

$k$2

and the selected alternative is

$k$3

The fixed-budget robust OCBA problem is

$k$4

Here PICS is the probability of incorrect selection, namely $k$5 under the stated ordering (Wan et al., 2024).

3. Additive upper bound and dimensional reduction

The defining technical feature of additive OCBA is an additive upper bound on PICS. Starting from the exact robust PICS event, the robust paper derives

$k$6

This decomposition uses only comparisons involving scenario $k$7, the worst-case scenario of the true best alternative. It contains $k$8 pairwise terms: $k$9 comparisons against the worst-case scenarios of non-best alternatives, and $i$0 comparisons against the non-worst scenarios of the best alternative (Wan et al., 2024).

This bound is “additive” in a literal sense: the surrogate objective is a sum of pairwise mis-ordering probabilities rather than a coupling across all $i$1 scenarios. The same paper contrasts this with the multiplicative upper bound

$i$2

which involves $i$3 terms (Wan et al., 2024).

Under normal outputs $i$4, the additive surrogate becomes

$i$5

where

$i$6

The paper then re-indexes the critical scenarios into a set $i$7, with $i$8 corresponding to $i$9, and rewrites the objective as

$X_{i,l} \sim \mathcal{N}(\mu_i,\sigma_i^2)$0

This is exactly the classical OCBA problem for a single-layer ranking-and-selection problem with $X_{i,l} \sim \mathcal{N}(\mu_i,\sigma_i^2)$1 alternatives and alternative $X_{i,l} \sim \mathcal{N}(\mu_i,\sigma_i^2)$2 as the best (Wan et al., 2024).

The structural consequence is sparse asymptotic allocation. The robust paper proves that

$X_{i,l} \sim \mathcal{N}(\mu_i,\sigma_i^2)$3

so asymptotically no budget is spent on non-critical scenarios. Only all scenarios of the best alternative and the worst-case scenario of each non-best alternative remain active in the limiting allocation (Wan et al., 2024).

4. Asymptotically optimal allocation rules

Because the reduced problem is a classical OCBA problem, the asymptotic optimal allocation follows the usual variance-to-gap structure. In the re-indexed form, as $X_{i,l} \sim \mathcal{N}(\mu_i,\sigma_i^2)$4, the optimal solution satisfies

$X_{i,l} \sim \mathcal{N}(\mu_i,\sigma_i^2)$5

and

$X_{i,l} \sim \mathcal{N}(\mu_i,\sigma_i^2)$6

In the original robust indices, the full asymptotic solution is

$X_{i,l} \sim \mathcal{N}(\mu_i,\sigma_i^2)$7

together with the ratio conditions

$X_{i,l} \sim \mathcal{N}(\mu_i,\sigma_i^2)$8

$X_{i,l} \sim \mathcal{N}(\mu_i,\sigma_i^2)$9

$b = \arg\max_i \mu_i$0

and

$b = \arg\max_i \mu_i$1

These equations imply that larger variance and smaller mean gap receive more budget, exactly as in classical OCBA, but now only over the critical scenarios singled out by the additive bound (Wan et al., 2024).

This reduced-dimension inheritance from classical OCBA is consistent with the broader OCBA theory. In the Gaussian finite-arm setting, Chen-type OCBA optimality conditions can be written as

$b = \arg\max_i \mu_i$2

with the balancing equation

$b = \arg\max_i \mu_i$3

while the large-deviations OCBA system is

$b = \arg\max_i \mu_i$4

$b = \arg\max_i \mu_i$5

The robust additive formulas are thus classical OCBA equations applied on a transformed index set defined by the additive PICS decomposition (Li et al., 2022).

A plausible implication is that the distinctive contribution of additive OCBA is not a new asymptotic algebra, but a new identification of which comparisons should enter the algebra.

5. Sequential procedures and additive implementation

The robust additive theory is instantiated algorithmically by AR-OCBA. The meta procedure begins with $b = \arg\max_i \mu_i$6 observations per scenario, computes sample means and sample variances, repeatedly identifies for each alternative its current worst-case scenario

$b = \arg\max_i \mu_i$7

then identifies the current best alternative

$b = \arg\max_i \mu_i$8

forms the reduced scenario set $b = \arg\max_i \mu_i$9 of size $n$0, computes updated OCBA allocations within $n$1, and allocates the next batch $n$2 accordingly (Wan et al., 2024).

The concrete AR-OCBA procedure uses a proportional stage-wise allocation rule. If $n$3 is the new target allocation for scenario $n$4, the stage-$n$5 sample increment is

$n$6

The paper contrasts this with a most-starving rule that allocates all $n$7 samples to the scenario with the largest deficit. The proportional rule is presented as the distinguishing feature of AR-OCBA among additive robust procedures (Wan et al., 2024).

This stage-wise structure aligns with the additive interpretation of OCBA developed in the convergence-rate analysis of OCBA-1 and OCBA-2. OCBA-1 computes estimated OCBA weights, forms target counts $n$8, and adds a batch $n$9 to the design with the largest deficit

$\alpha_i = n_i/n$0

whereas OCBA-2 uses an index rule derived from the large-deviations optimality conditions (Li et al., 2022). Both are sequential OCBA algorithms in which budget is accumulated incrementally. The convergence-rate paper proves that

$\alpha_i = n_i/n$1

for OCBA-1 and

$\alpha_i = n_i/n$2

for OCBA-2, so the additive sample-by-sample or batch-by-batch implementation recovers the static OCBA target asymptotically (Li et al., 2022).

The stochastic-time extension OCBAS provides another additive implementation motif. There the decision variables are times $\alpha_i = n_i/n$3, not sample counts $\alpha_i = n_i/n$4, and the number of completed replications is random. The key asymptotic approximation is

$\alpha_i = n_i/n$5

which yields the OCBAS allocation conditions

$\alpha_i = n_i/n$6

and

$\alpha_i = n_i/n$7

This suggests that additive OCBA can also be interpreted as OCBA under additively accumulated non-replication budgets, provided estimator variance can be written in effective-budget form (Jia, 2012).

6. Empirical behavior, performance measures, and scope

The robust additive paper reports a comprehensive numerical study comparing AR-OCBA, AR-OCBA with the most-starving rule, equal allocation, and R-OCBA. Under monotone mean configurations and equal, increasing, or decreasing variance configurations, AR-OCBA has consistently higher PCS than R-OCBA across all budgets, and the PCS gap widens as the number of alternatives $\alpha_i = n_i/n$8 or the number of distributions $\alpha_i = n_i/n$9 increases (Wan et al., 2024). In the stage-wise allocation study with $\sigma_i^2/n_i$0 and $\sigma_i^2/n_i$1, proportional AR-OCBA significantly outperforms AR-OCBA with the most-starving rule and equal allocation; in some settings the most-starving version performs worse than equal allocation (Wan et al., 2024).

The same study reports that PCS is fairly insensitive to the initialization $\sigma_i^2/n_i$2 in a broad range and also insensitive to $\sigma_i^2/n_i$3 in the tested range. In a $\sigma_i^2/n_i$4 example, AR-OCBA allocates almost all budget to only $\sigma_i^2/n_i$5 scenarios, while the other scenarios remain essentially at their initial $\sigma_i^2/n_i$6, matching the sparse allocation predicted by Theorem 3.3 (Wan et al., 2024).

The sequential OCBA analysis provides complementary performance characterizations. For OCBA-1 and OCBA-2 with known variances, probability of false selection and expected opportunity cost decay exponentially: $\sigma_i^2/n_i$7 for OCBA-1, and similarly with $\sigma_i^2/n_i$8 for OCBA-2, while cumulative regret grows linearly: $\sigma_i^2/n_i$9 or with $k+m-1$00 for OCBA-2 (Li et al., 2022). Modified UM variants replace this with logarithmic cumulative regret but only polynomial decay in PFS and EOC (Li et al., 2022). This suggests that additive OCBA can target different asymptotic criteria depending on whether the primary objective is final-selection quality or cumulative sampling efficiency.

The stochastic-time study reports that OCBAS is asymptotically optimal, that the variance of simulation time does not appear in the asymptotic variance formula, and that numerical results show OCBA for deterministic simulation time is robust even when the simulation time is stochastic (Jia, 2012). The paper also states that, using Lemma 3, it is possible to extend OCBAS to handle multiple objective functions, simulation-based constraints, opportunity cost, and complexity preferences, following its according extensions in OCBA (Jia, 2012). A plausible implication is that additive OCBA may generalize beyond robust R&S whenever structured budgets or structured error events admit an additive reduction to an effective OCBA problem.

The current scope remains defined by the assumptions used in the source formulations. In robust additive OCBA these include a finite ambiguity set, the mean-configuration assumption, normality of scenario outputs, and large-budget asymptotics (Wan et al., 2024). The paper identifies as future directions a rigorous consistency analysis of AR-OCBA, a theoretical characterization of the additive structure in robust R&S, and extensions to infinite or uncountable ambiguity sets (Wan et al., 2024). These limitations indicate that additive OCBA is presently best viewed as an asymptotic OCBA framework whose principal technical innovation is the replacement of the full robust comparison structure by an additive critical-comparison structure.