Towards Automated Knowledge Transfer in Evolutionary Multitasking via Large Language Models

1. Problem Statement

Evolutionary Multi-Task Optimization (EMTO) is a paradigm that solves multiple optimization tasks simultaneously by transferring knowledge between them to improve search efficiency. However, the core component of EMTO, the Knowledge Transfer Method (KTM), faces significant challenges:

• Scenario Dependency: Existing KTMs (e.g., genetic crossover, subspace mapping, neural networks) are often tailored to specific problem settings. A method effective for high-similarity tasks may fail or cause "negative transfer" in low-similarity scenarios.
• Manual Design Bottleneck: Designing robust KTMs requires substantial domain expertise and manual customization, limiting scalability and adaptability across diverse optimization landscapes.
• Limitations of Current LLM Applications: While Large Language Models (LLMs) have been used as direct search operators, they struggle with numerical computation and high-dimensional spaces. Conversely, using LLMs to generate entire solvers often lacks specific guidance for the complex, flexible components required in multitasking (like knowledge transfer mechanisms).

The paper addresses the open problem of automating the design of KTMs to create scenario-adaptive transfer strategies without heavy human intervention.

2. Methodology: Self-guided Knowledge Transfer Design (SKTD)

The authors propose SKTD, a framework that leverages LLMs to autonomously generate, evolve, and refine KTMs. The framework operates as a cyclic search process with three tightly integrated components:

A. Automated KTM Design (The Search Engine)

SKTD treats KTM design as an automated search problem. It maintains a population of candidate KTMs (Python functions) and evolves them using three LLM-driven operators:

1. Initialize: Generates an initial population of KTMs based on prompts describing the problem and existing literature.
2. Recombine: Synthesizes new KTMs by combining useful sub-modules from high-performing parent KTMs.
3. Discover: Promotes innovation by instructing the LLM to create novel KTMs that differ significantly from existing ones, expanding the search space.

• Repair & Refinement: A dedicated module ensures all LLM-generated code is syntactically correct, executable, and adheres to the required input/output interface before evaluation.

B. Self-guided Knowledge Extraction Mechanism

To prevent the search from being purely random, SKTD extracts structured design insights from the search trajectory to guide future generations. This occurs at two levels:

1. Individual-level: Analyzes pairs where an offspring outperforms its parent. The LLM performs a contrastive analysis to identify specific algorithmic improvements (fine-grained insights).
2. Population-level: Compares high-performing KTMs from the current generation against underperforming ones from previous generations to extract general evolutionary trends and structural regularities (coarse-grained insights).

C. Knowledge-Guided Improvement (The "Improve" Operator)

The extracted insights are stored in a semantic knowledge base. The Improve Operator utilizes this knowledge to refine existing KTMs:

• It selects a candidate KTM and retrieves relevant insights from the knowledge base based on semantic similarity.
• It prompts the LLM to apply these specific tips to enhance the KTM's performance while preserving its core functionality.
• This creates a feedback loop where the system learns from its own search history to progressively refine transfer strategies.

3. Key Contributions

1. First Automated KTM Generation: This work represents the first attempt to automate the generation of Knowledge Transfer Methods for EMTO using LLMs, moving beyond manual design or simple program search.
2. Integrated SKTD Framework: The authors developed a novel paradigm that combines LLM-driven program synthesis with a self-guided knowledge extraction mechanism, enabling the systematic construction of scenario-adaptive KTMs.
3. Self-Guided Knowledge Utilization: A unique mechanism is introduced to distill both fine-grained (individual) and coarse-grained (population) design insights from search trajectories, using them to explicitly guide the LLM in refining transfer strategies.
4. Robustness Across Similarity Levels: The method demonstrates the ability to autonomously construct effective KTMs that perform well across varying degrees of task similarity (high, medium, and low), a known weakness of traditional hand-crafted methods.

4. Experimental Results

The authors evaluated SKTD on nine multi-task optimization benchmarks with varying task similarities (using the CEC 2022 test suite) and compared it against:

• Hand-crafted KTMs: Vertical Crossover (VC), Orthogonal Transfer (OT), Autoencoding (AE), and Affine Transformation (AF).
• State-of-the-art LLM method: EoH (Evolution of Heuristics).

Key Findings:

• Superior Performance: SKTD achieved the best normalized fitness values in 5 out of 9 scenarios, outperforming both hand-crafted methods and the EoH baseline.
• Handling Low Similarity: Hand-crafted methods often failed (normalized fitness > 1.0) in low-similarity scenarios due to negative transfer. SKTD consistently maintained performance below 0.7 across all scenarios, demonstrating superior generalization.
• Ablation Study: Removing the "Improve" operator (SKTD-NI) resulted in significantly worse performance, particularly in medium-to-low similarity scenarios. This confirms that the self-guided knowledge extraction is critical for success.
• Search Dynamics: Analysis showed that while initialization seeds the search, the Recombine and Improve operators drive the majority of performance gains. The "Improve" operator, though used less frequently, was often responsible for the best-performing KTMs, either directly or by injecting knowledge into the lineage of future candidates.
• Knowledge Evolution: Keyword analysis of the extracted knowledge revealed that the system adapts its focus over time (e.g., shifting from boundary handling to clustering strategies) and varies its focus based on the specific problem scenario.

5. Significance

This research marks a significant step toward autonomous algorithm design in evolutionary computation. By automating the creation of knowledge transfer mechanisms, SKTD reduces the reliance on human domain expertise and overcomes the rigidity of fixed transfer strategies.

• Practical Impact: It offers a scalable solution for complex real-world optimization problems where task characteristics vary dynamically.
• Theoretical Advancement: It demonstrates that LLMs can effectively act not just as code generators, but as meta-learners capable of synthesizing high-level algorithmic components by learning from their own evolutionary search history.
• Future Direction: The work opens new avenues for integrating advanced AI techniques into evolutionary multitasking, potentially extending to real-world applications in engineering, robotics, and finance.