Particle algorithms for optimization on binary spaces

We discuss a unified approach to stochastic optimization of pseudo-Boolean objective functions based on particle methods, including the cross-entropy method and simulated annealing as special cases. We point out the need for auxiliary sampling distributions, that is parametric families on binary spaces, which are able to reproduce complex dependency structures, and illustrate their usefulness in our numerical experiments. We provide numerical evidence that particle-driven optimization algorithms based on parametric families yield superior results on strongly multi-modal optimization problems while local search heuristics outperform them on easier problems.

💡 Research Summary

The paper presents a unified particle‑based framework for global optimization of pseudo‑Boolean functions defined on binary vectors. Starting from the observation that many stochastic optimization methods can be interpreted as sequential importance sampling on a family of probability distributions, the authors introduce two canonical families: a tempered family πβ(γ)∝exp(β f(γ)) and a level‑set family πβ(γ)∝1_{L⁺β}(γ)/|L⁺β|, where L⁺β contains all binary vectors whose objective value lies within 1/β of the global optimum. As β→∞ both families concentrate on the set of global maximizers, thus providing a probabilistic bridge between simulated annealing (tempered family) and the cross‑entropy method (level‑set family).

To exploit these families in practice the authors adopt a Sequential Monte Carlo (SMC) algorithm. An initial particle cloud is drawn uniformly from the binary hypercube. At each iteration the algorithm (i) updates importance weights to target the next distribution in the β‑schedule, (ii) monitors the effective sample size (ESS) and chooses the step length α so that ESS is reduced by a fixed proportion (β≈0.9), (iii) performs systematic resampling to eliminate low‑weight particles, and (iv) applies a Markov transition kernel to diversify the particle set. The transition kernel can be a symmetric local move (flipping a random subset of bits) or, more importantly, an adaptive independent kernel qθ(·) whose parameters θ are fitted to the current particle approximation (e.g., by maximum likelihood or KL‑minimisation).

A central contribution is the systematic study of three parametric families for the independent proposal distribution on binary spaces: (1) independent Bernoulli product models, (2) multivariate Bernoulli (multinomial logit) models that capture pairwise correlations, and (3) Ising‑like graphical models that encode arbitrary pairwise interactions. By fitting these families to the particle cloud, the proposal distribution can closely approximate the target πβ, leading to high acceptance rates and rapid mixing of the independent kernel. The authors also introduce a particle‑diversity measure ζn and stop the move steps when ζn stabilises, thereby preventing unnecessary computation.

Experimental evaluation focuses on unconstrained quadratic binary optimization (QUBO) instances, which are representative of many combinatorial problems. The authors compare four algorithmic configurations: (a) SMC with each of the three parametric families, (b) the classic cross‑entropy method, (c) simulated annealing, and (d) simple multi‑restart local search heuristics such as k‑opt. Two classes of problem instances are considered: (i) strongly multimodal, high‑dimensional instances with many deep local optima, and (ii) easier, low‑dimensional or nearly unimodal instances. Results show that on the multimodal class, SMC equipped with a rich parametric family (especially the Ising‑like model) consistently finds solutions closer to the global optimum than the other methods, confirming the advantage of a globally‑aware particle exploration. In contrast, on the easier class, local heuristics achieve comparable or better performance with far less computational effort, highlighting the trade‑off between algorithmic sophistication and problem difficulty.

The paper concludes that particle‑driven global optimization, when combined with expressive parametric proposal families and adaptive SMC machinery, offers a powerful tool for binary optimization problems that exhibit strong multimodality. However, the approach incurs higher computational overhead and requires careful tuning of the parametric model and β‑schedule. Future work is suggested in three directions: (1) designing more expressive yet tractable parametric families (e.g., deep Bayesian networks), (2) developing faster adaptive independent kernels, and (3) scaling the methodology through parallel and GPU implementations. Overall, the work bridges the gap between classical meta‑heuristics and modern sequential Monte Carlo techniques, providing a solid theoretical and empirical foundation for advanced binary optimization.