A Scalable Test Problem Generator for Sequential Transfer Optimization.

Area of Science:

• Machine Learning
• Optimization Algorithms
• Artificial Intelligence

Background:

• Sequential Transfer Optimization (STO) is gaining interest, but lacks standardized benchmarks.
• Existing STO test problems are often manually configured, limiting scalability and leading to biased algorithm performance.
• A systematic comparison of STO algorithms is hindered by the absence of comprehensive, scalable evaluation tools.

Purpose of the Study:

• To introduce a novel benchmark suite for systematically evaluating Sequential Transfer Optimization (STO) algorithms.
• To address the limitations of existing STO test problems, including manual configuration and lack of scalability.
• To provide a platform for uncovering nuanced algorithm behaviors through diverse and customizable task relationships.

Main Methods:

• Introduced four concepts for characterizing STO problems (STOPs).
• Defined 'similarity distribution' to quantitatively assess relationships between source and target task optimal solutions.
• Developed a scalable problem generator with a novel inverse strategy for customizing similarity distributions.

Main Results:

• Created a benchmark suite of 12 STOPs with customized similarity relationships.
• The new benchmark revealed findings like biased transferability representation and performance improvements unrelated to search experience, previously undetected.
• Demonstrated the generator's scalability and ability to capture diverse, real-world similarity relationships.

Conclusions:

• The developed benchmark suite offers a robust platform for evaluating STO algorithms.
• The novel problem generator and similarity distribution metric enable more reliable and generalizable STO research.
• This work facilitates deeper understanding of STO algorithm behavior and performance across varied task scenarios.