An Evolutionary Algorithm with Probabilistic Annealing for Large-scale Sparse Multi-objective Optimization
Novel 'probabilistic annealing' technique balances exploration vs. exploitation in high-dimensional search spaces.
A research team from multiple institutions, led by Shuai Shao, has published a preprint paper introducing a novel evolutionary algorithm designed to tackle Large-scale Sparse Multi-objective Optimization Problems (LSMOPs). These problems, common in fields like adversarial AI attacks and critical node detection, involve finding optimal solutions where only a few decision variables are non-zero within an extremely high-dimensional search space. The core challenge is the intense conflict between exploring the vast search area and exploiting promising regions, which overwhelms existing Multi-Objective Evolutionary Algorithms (MOEAs).
The proposed solution, an Evolutionary Algorithm with Probabilistic Annealing, innovates by using two distinct probability vectors to guide the search. A convergence-oriented vector with low entropy ensures stable exploitation of good solutions, while an 'annealed' probability vector starts with high entropy for broad exploration and gradually reduces it to focus on local refinement. This dynamic integration creates an adaptive balance, allowing the algorithm to efficiently identify the critical non-zero variables. Experimental results on benchmark problems demonstrate that this method outperforms current state-of-the-art evolutionary algorithms in both convergence speed and the diversity of solutions found.
- Solves Large-scale Sparse Multi-objective Optimization Problems (LSMOPs) where optimal solutions have very few non-zero variables.
- Uses two probability vectors with distinct entropy: one for stable exploitation and an annealed one for adaptive exploration-to-refinement transition.
- Outperforms existing state-of-the-art evolutionary algorithms in convergence and diversity on benchmarks and real-world applications like adversarial attacks.
Why It Matters
Enables more efficient AI model optimization, adversarial robustness testing, and complex system design where traditional methods fail.