ROMC: Robust Optimization Monte Carlo

• ROMC is a global optimization method that transforms the objective function into a sequence of annealed densities concentrating on the global optimum.
• It employs sequential importance sampling combined with posterior exploration, including EM updates and componentwise Metropolis moves, to adaptively refine sampling.
• Empirical results show that ROMC robustly handles multimodal and high-dimensional optimization challenges, outperforming several standard techniques.

Robust Optimization Monte Carlo (ROMC) is not mentioned in the original source, but the "Posterior Exploration based Sequential Monte Carlo for Global Optimization" (PE-SMC) algorithm provides a rigorous, annealing-based Sequential Monte Carlo (SMC) approach to global optimization. PE-SMC transforms the objective function into a sequence of annealed target densities, applies adaptive importance sampling, and uses a posterior exploration subroutine to iteratively refine the proposal distribution and annealing schedule, concentrating sampling efforts on regions likely to contain the global optimum (Liu, 2015).

1. Problem Formulation and Annealed Target Sequence

The method addresses global maximization of a non-negative, bounded objective $f(x)$ over $x \in \mathcal{X} \subset \mathbb{R}^d$, seeking $f^* = \max_x f(x)$. It constructs a sequence of annealed target densities:

$$\pi_k(x) \propto T_k(x) = [f(x)]^{\lambda_k}, \qquad 0 < \lambda_1 < \lambda_2 < \cdots \to \infty,$$

where $\lambda_k$ is the inverse-temperature parameter. As $\lambda_k \to \infty$, $T_k(x)$ places nearly all its probability mass on the set $\{x : f(x) = f^*\}$, focusing samples on the global maximizers.

The challenge is to track $\pi_k$ sequentially and adapt $\lambda_k$ such that each $x \in \mathcal{X} \subset \mathbb{R}^d$0 is not "too far" from $x \in \mathcal{X} \subset \mathbb{R}^d$1, enabling efficient exploration while concentrating on optima.

2. Sequential Importance Sampling Framework

At each iteration $x \in \mathcal{X} \subset \mathbb{R}^d$2, the algorithm uses an importance sampling (IS) density $x \in \mathcal{X} \subset \mathbb{R}^d$3 parameterized by $x \in \mathcal{X} \subset \mathbb{R}^d$4. $x \in \mathcal{X} \subset \mathbb{R}^d$5 i.i.d. particles are sampled:

$x \in \mathcal{X} \subset \mathbb{R}^d$6

with importance weights

$x \in \mathcal{X} \subset \mathbb{R}^d$7

normalized as

$x \in \mathcal{X} \subset \mathbb{R}^d$8

Particles may be resampled to yield an unweighted cloud approximating $x \in \mathcal{X} \subset \mathbb{R}^d$9. In PE-SMC, most resampling and propagation is handled within the posterior exploration step.

3. Posterior Exploration (PE) Subroutine

The posterior exploration (PE) procedure is central, with three objectives:

• Explore high-probability regions of $f^* = \max_x f(x)$0 possibly underrepresented by the proposal.
• Adaptively refine $f^* = \max_x f(x)$1 to approximate $f^* = \max_x f(x)$2.
• Provide a mechanism for selecting $f^* = \max_x f(x)$3 for the next target.

PE employs four key operators:

A. Importance Sampling: Draw samples from $f^* = \max_x f(x)$4 and compute weights to approximate $f^* = \max_x f(x)$5.

B. Componentwise Metropolis Move: For each particle and each coordinate, propose a Gaussian perturbation and accept it with probability $f^* = \max_x f(x)$6, updating the particle if accepted.

C. EM Update of IS Density: $f^* = \max_x f(x)$7 is a mixture of $f^* = \max_x f(x)$8 multivariate Student's-$f^* = \max_x f(x)$9 components:

$$\pi_k(x) \propto T_k(x) = [f(x)]^{\lambda_k}, \qquad 0 < \lambda_1 < \lambda_2 < \cdots \to \infty,$$0

Given the weighted samples, a single EM step updates component weights $$\pi_k(x) \propto T_k(x) = [f(x)]^{\lambda_k}, \qquad 0 < \lambda_1 < \lambda_2 < \cdots \to \infty,$$1, means $$\pi_k(x) \propto T_k(x) = [f(x)]^{\lambda_k}, \qquad 0 < \lambda_1 < \lambda_2 < \cdots \to \infty,$$2, and covariances $$\pi_k(x) \propto T_k(x) = [f(x)]^{\lambda_k}, \qquad 0 < \lambda_1 < \lambda_2 < \cdots \to \infty,$$3.

D. Adding New Components: If the effective sample size (ESS) $$\pi_k(x) \propto T_k(x) = [f(x)]^{\lambda_k}, \qquad 0 < \lambda_1 < \lambda_2 < \cdots \to \infty,$$4 falls below a threshold, the highest-weight particle is identified, and a new Student's-$$\pi_k(x) \propto T_k(x) = [f(x)]^{\lambda_k}, \qquad 0 < \lambda_1 < \lambda_2 < \cdots \to \infty,$$5 component is added at its location. This process iterates, with new samples merged and EM/IS steps repeated, until ESS exceeds the threshold.

PE's steps are summarized as follows:

Step	Description	Stopping Condition
Importance Sampling	Draw samples from $$\pi_k(x) \propto T_k(x) = [f(x)]^{\lambda_k}, \qquad 0 < \lambda_1 < \lambda_2 < \cdots \to \infty,$$6, compute/update weights
Metropolis Move	Componentwise update of each particle using local perturbations
EM Update	Re-estimate mixture parameters using weighted samples
New Component Addition	Add components at high-weight particles as needed, re-sample/merge to boost ESS	Stop when $$\pi_k(x) \propto T_k(x) = [f(x)]^{\lambda_k}, \qquad 0 < \lambda_1 < \lambda_2 < \cdots \to \infty,$$7

4. Adaptive Annealing Schedule

The annealing schedule is selected online after each PE iteration. For any $$\pi_k(x) \propto T_k(x) = [f(x)]^{\lambda_k}, \qquad 0 < \lambda_1 < \lambda_2 < \cdots \to \infty,$$8, a hypothetical $$\pi_k(x) \propto T_k(x) = [f(x)]^{\lambda_k}, \qquad 0 < \lambda_1 < \lambda_2 < \cdots \to \infty,$$9 is computed by reweighting current samples. The next annealing parameter $\lambda_k$0 is chosen to solve:

$\lambda_k$1

where $\lambda_k$2 and $\lambda_k$3 is a user constant, empirically $\lambda_k$4. This ensures sufficient continuity between successive targets to maintain IS efficiency.

5. Theoretical Properties and Weak Convergence

The algorithm achieves several guarantees:

• As $\lambda_k$5, $\lambda_k$6 converges weakly to a uniform distribution over the set of global optima $\lambda_k$7 (Proposition 1).
• Each PE iteration (SIS + EM + Metropolis + component-addition) increases the effective sample size and drives $\lambda_k$8 closer to $\lambda_k$9.
• The particle approximation tracks $\lambda_k \to \infty$0 closely, so as $\lambda_k \to \infty$1 ($\lambda_k \to \infty$2), particles concentrate around the global maximizers.
• Formal weak-convergence proofs are provided.

6. Empirical Benchmarking and Comparative Results

PE-SMC was evaluated on 17 multimodal benchmark functions (dimensions 2 to 20) and compared to:

• Particle-filter-optimization (PFO)
• SMC-simulated annealing (SMC-SA)
• Trelea’s particle swarm optimization (PSO)

Performance measures (average ± std. over 100 runs) are reported. For example, on the 2D Rastrigin problem (max=200):

Algorithm	2D Rastrigin	20D Rastrigin
SMC-SA	199.9990 ± 1.2×10⁻³	196.3677 ± 0.533
PE-SMC	199.9999 ± 1.1×10⁻⁶	199.9228 ± 0.0113

Over 23 test cases, PE-SMC achieved the best result in 10, the second-best in 8, and was never the worst. SMC-SA and PFO were outperformed by PE-SMC, especially in higher dimensions. PSO achieved best results in 12 cases but was the worst in 3.

7. Applications and Significance

The methodology provides a flexible approach for continuous, multimodal global optimization tasks, especially those where the objective is evaluable but derivatives are unavailable or uninformative. The PE subroutine enables the algorithm to adaptively address sample impoverishment and modal structure, making it well-suited for high-dimensional, rugged landscapes. The adaptively constructed annealing schedule and proposal distribution distinguish this approach from standard SMC-based optimization frameworks.

A plausible implication is that algorithms based on this framework, with its systematic annealing and adaptive proposal refinement via posterior exploration, are especially robust to multimodality and challenging sample allocation problems encountered in global optimization of complex, continuous functions (Liu, 2015).