Transferable Adaptive Differential Evolution for Many-Task Optimization.

Area of Science:

• Evolutionary Computation
• Optimization Algorithms
• Machine Learning

Background:

• Evolutionary multitask optimization (EMTO) addresses many-task optimization problems (MaTOPs).
• Existing EMTO methods rely on population distribution similarity for knowledge transfer (KT), which can be ineffective when task optima differ significantly.
• A novel similarity measure is needed to improve KT in MaTOPs.

Purpose of the Study:

• To propose a new definition of task similarity: shift invariance.
• To develop a novel EMTO algorithm, transferable adaptive differential evolution (TRADE), that leverages shift invariance for improved KT.
• To evaluate TRADE's performance against state-of-the-art algorithms on MaTOP benchmarks.

Main Methods:

• Introduced the concept of shift invariance, defining task similarity after linear transformations in search and objective spaces.
• Developed a two-stage TRADE algorithm: task representation and grouping based on shift invariance, followed by adaptive parameter transfer.
• Utilized a task representation strategy to embed evolution information into task vectors for grouping.

Main Results:

• The proposed TRADE algorithm successfully groups shift-invariant tasks.
• TRADE adaptively transfers successful evolution experiences and parameters among similar tasks within groups.
• Experiments on MaTOP benchmarks and a real-world application demonstrated TRADE's superiority over existing EMTO and single-task algorithms.

Conclusions:

• Shift invariance offers a more effective similarity measure for EMTO than traditional methods.
• The TRADE algorithm demonstrates significant improvements in solving MaTOPs by exploiting shift-invariant task relationships.
• TRADE provides a robust and adaptive approach for knowledge transfer in complex optimization scenarios.