[Paper] Adaptive Surrogate-Based Strategy for Accelerating Convergence Speed when Solving Expensive Unconstrained Multi-Objective Optimisation Problems

Source: arXiv - 2601.21885v1

Overview

The paper introduces an adaptive surrogate‑based accelerator that plugs into existing multi‑objective evolutionary algorithms (MOEAs) to dramatically cut the number of expensive fitness evaluations needed during the early stages of optimization. By swapping costly true evaluations for fast, learned approximations, the method makes MOEAs practical for real‑world, compute‑intensive problems such as large‑scale environmental modeling.

Key Contributions

• Two‑loop architecture: an outer loop runs a conventional MOEA (e.g., NSGA‑II, MOEA/D) with true evaluations, while an inner loop repeatedly trains and queries surrogate models to guide the search.
• Adaptive surrogate selection: the system automatically chooses among three surrogate families—Gaussian Process Regression (GPR), 1‑D Convolutional Neural Networks (CNN‑1D), and Random Forest Regression (RFR)—based on current data quality and prediction confidence.
• Early‑stage convergence boost: empirical results on 31 benchmark suites and a North Sea fish‑abundance case study show up to 3–5× reduction in the number of true evaluations required to reach comparable Pareto front quality.
• Framework‑agnostic integration: the accelerator can be wrapped around any off‑the‑shelf MOEA without modifying its internal operators, preserving algorithmic guarantees while adding speed.

Methodology

1. Baseline MOEA (host) – Runs as usual, generating candidate solutions and occasionally invoking the true, expensive fitness function (e.g., a high‑fidelity simulation).
2. Data collection – Each true evaluation is stored in a growing dataset of decision vectors ↔ objective values.
3. Surrogate training – After a predefined “warm‑up” period, the system trains three surrogate models on the collected data:
   • GPR for smooth, low‑dimensional problems (provides uncertainty estimates).
   • CNN‑1D for problems where the decision vector exhibits spatial or temporal structure (e.g., time‑series control parameters).
   • RFR for high‑dimensional, noisy landscapes.
4. Adaptive selection – The accelerator evaluates each surrogate’s cross‑validation error and, when available, its predictive variance. The model with the best trade‑off between accuracy and confidence is chosen for the next inner‑loop iteration.
5. Inner loop (Accelerator) – The selected surrogate predicts fitness for a batch of newly generated candidates. Those with promising surrogate scores are promoted to the outer loop for true evaluation, while the rest are discarded or kept for later refinement.
6. Iterative refinement – As more true evaluations accumulate, surrogates are retrained, gradually improving their fidelity and allowing the outer MOEA to converge faster.

The whole pipeline is lightweight: surrogate training is performed on the CPU (or a modest GPU for the CNN), and the inner loop can evaluate thousands of candidates per second, far outpacing the original simulation.

Results & Findings

• Pareto front quality (measured by Inverted Generational Distance and Hypervolume) was statistically indistinguishable from the baseline after the early‑phase speed‑up.
• Model selection dynamics: GPR dominated early on (when data were scarce), RFR took over as dimensionality grew, and CNN‑1D became favorable once the decision vectors exhibited clear sequential patterns.
• Robustness: The approach maintained its advantage across different MOEA variants (NSGA‑II, MOEA/D) and problem scales (2–10 objectives, 10–200 decision variables).

Practical Implications

• Cost‑effective R&D: Companies that run costly CFD, climate, or bio‑simulation loops can now embed the accelerator into their existing evolutionary pipelines, cutting compute budgets by up to 80 % during prototyping.
• Faster time‑to‑market: Product‑design teams can iterate on multi‑objective trade‑offs (e.g., weight vs. strength vs. cost) in hours rather than days, enabling more aggressive exploration of the design space.
• Edge‑compatible optimization: Because surrogate inference is cheap, the inner loop can be off‑loaded to edge devices or low‑power servers, allowing distributed optimization across a fleet of IoT sensors or autonomous agents.
• Plug‑and‑play tooling: The authors released a Python package (surrogate‑moea) that wraps around DEAP, pymoo, or any custom MOEA, requiring only a callable fitness function. Developers can thus adopt the technique with minimal code changes.

Limitations & Future Work

• Surrogate reliability: The method hinges on surrogate accuracy; in highly chaotic or discontinuous fitness landscapes, prediction errors may misguide the search, necessitating more frequent true evaluations.
• Scalability of GPR: Gaussian Processes scale cubically with the number of training points, so for problems requiring > 10 000 true evaluations the authors suggest sparse GP variants or switching to RFR.
• Domain‑specific feature engineering: The CNN‑1D surrogate works best when the decision vector has an inherent ordering; other problem types may need custom architectures.
• Future directions: The authors plan to explore active‑learning strategies for smarter selection of which surrogate‑predicted candidates to promote, and to integrate reinforcement‑learning controllers that dynamically adjust the inner‑loop budget based on convergence signals.

Authors

• Tiwonge Msulira Banda
• Alexandru‑Ciprian Zăvoianu

Paper Information

• arXiv ID: 2601.21885v1
• Categories: cs.NE
• Published: January 29, 2026
• PDF: Download PDF