Online Automated Algorithm Design with Large Language Models
OnDesign leverages online LLMs for algorithm design, surpassing traditional optimizers.
Key Findings
Methodology
OnDesign is a multi-agent framework integrating state analysis, agent deliberation, and SAG evolution modules. Each plays a crucial role in dynamically generating algorithms during optimization.
Key Results
- On BBOB 30D, OnDesign achieved an average rank of 2.02, outperforming FunBO's 5.10.
- On CEC2026 30D, OnDesign's average rank was 2.84, better than JES's 4.66.
- In all experiments, OnDesign consistently ranked best across 17 configurations.
Significance
This study significantly enhances optimization efficiency by integrating algorithm design with target optimization, especially in dynamic environments, offering a new paradigm for academia and industry.
Technical Contribution
OnDesign breaks the limitations of traditional optimizers by treating algorithms as state-dependent decision variables, offering new theoretical guarantees and engineering possibilities.
Novelty
First to use LLMs for online algorithm design, distinct from previous offline methods, significantly enhancing optimization flexibility.
Limitations
- High computational cost in high-dimensional problems.
- Requires substantial initial data for model training.
Future Work
Future research could explore applications in more scenarios, such as real-time systems, and further optimize computational efficiency.
AI Executive Summary
Traditional algorithm design often separates design from optimization, leading to poor performance in dynamic environments. OnDesign introduces a new paradigm by integrating algorithm design with target optimization using online LLMs.
The core of OnDesign lies in its multi-agent framework, comprising state analysis, agent deliberation, and SAG evolution modules. These modules work together to dynamically generate algorithms that adapt to the current optimization state.
Experimental results show that OnDesign performs exceptionally well across multiple benchmarks, especially in high-dimensional problems. It not only enhances optimization efficiency but also provides new directions for future research. However, computational costs and initial data requirements remain challenges to be addressed.
Deep Analysis
Background
Recent advances in LLMs have significantly impacted natural language processing, and their application in algorithm design is gaining attention. Traditional methods rely heavily on expert knowledge and struggle to adapt to dynamic optimization environments.
Core Problem
Existing algorithm design methods often separate design from optimization, leading to poor performance in dynamic environments. Integrating algorithm design with target optimization is a pressing issue.
Innovation
OnDesign introduces dynamic algorithm design using online LLMs. Its innovation lies in treating algorithms as state-dependent decision variables and employing a multi-agent framework for dynamic generation.
Methodology
- �� State Analysis: Guided by SAG to generate state reports.
- �� Agent Deliberation: Multiple agents propose designs, and an arbiter synthesizes the algorithm.
- �� SAG Evolution: Updates the analysis guideline based on feedback.
Experiments
Experiments were conducted on BBOB, CEC2020, CEC2022, and CEC2026 benchmarks, covering Bayesian optimization and evolutionary computation. Metrics include average rank and final objective value.
Results
OnDesign consistently outperformed baseline methods in all experiments, particularly in high-dimensional problems, with significant improvements in average rank.
Applications
OnDesign is suitable for scenarios requiring dynamic optimization, such as real-time systems and complex engineering problems. Its flexibility and efficiency offer broad industrial applications.
Limitations & Outlook
Despite its superior performance, OnDesign has high computational costs and requires substantial initial data. These issues need further exploration in future research.
Plain Language Accessible to non-experts
Imagine you're in a kitchen cooking. Traditional algorithm design is like following a fixed recipe, regardless of ingredient freshness or guest preferences. OnDesign is like a smart chef who adjusts the recipe in real-time based on ingredient changes and guest feedback to create the perfect dish.
ELI14 Explained like you're 14
Imagine playing a game where you always use the same strategy to win. That's traditional algorithm design. But OnDesign is like a super-smart player who changes strategies in real-time based on the game's changes and opponents' moves, ensuring you win every time! Isn't that cool?
Glossary
Large Language Model (LLM)
A model capable of understanding and generating natural language, commonly used in NLP tasks.
Used for algorithm design and code generation.
Automated Algorithm Design (AAD)
The process of automatically generating optimization strategies through algorithms.
Implemented online using LLMs.
Bayesian Optimization
A strategy for optimizing black-box functions, particularly suitable for expensive computations.
One of the experimental scenarios.
Evolutionary Computation
An optimization algorithm simulating natural evolution, suitable for complex problems.
Used for continuous and mixed-variable optimization.
State Analysis Guideline (SAG)
A framework guiding state analysis to generate optimization state reports.
Used in OnDesign for dynamic algorithm adjustment.
Open Questions Unanswered questions from this research
- 1 How to enhance OnDesign's efficiency without increasing computational costs?
- 2 How does OnDesign perform in more complex dynamic environments?
Applications
Immediate Applications
Real-time System Optimization
OnDesign can be used for dynamic optimization in real-time systems, improving response speed and efficiency.
Long-term Vision
Complex Engineering Problem Solving
In complex engineering problems, OnDesign can provide flexible optimization solutions, adapting to various constraints and requirements.
Abstract
Large language models (LLMs) enable automated algorithm design (AAD) through reasoning and code synthesis. However, most existing LLM-based AAD methods separate algorithm design from target optimization, deploying a fixed design even as the optimization state evolves. Conventional adaptive optimizers can respond to such changes, but their adjustments remain confined to predefined parameters, operators, or strategies. To address these limitations, we introduce online LLM-based AAD, a novel optimization paradigm that treats the algorithm itself as a state-dependent decision variable. At each stage, LLM agents synthesize an algorithm with new behavior logic from the current optimization state. Executing the generated algorithm advances the search and provides feedback for subsequent designs, coupling algorithm design with target optimization without requiring a separate offline algorithm pretraining stage. To implement this paradigm, we propose OnDesign, a multi-agent framework that reconciles competing design perspectives to synthesize executable algorithms and uses execution feedback to refine how runtime evidence is interpreted for subsequent designs. We evaluate OnDesign across two mainstream black-box optimization paradigms on three scenarios: Bayesian optimization, evolutionary continuous optimization, and evolutionary mixed-variable optimization. Extensive experiments on six benchmark suites and one engineering problem across multiple problem dimensions demonstrate superior overall performance over conventional optimizers and offline LLM-based AAD methods.