Domain Adaptation Multitask Optimization.

Area of Science:

• Optimization
• Evolutionary Computation
• Machine Learning

Background:

• Multitask optimization (MTO) leverages information from multiple tasks to improve individual task performance.
• Distinct tasks often have different solution distributions, posing challenges for intertask learning.
• Existing MTO methods can suffer from negative transfer due to distribution mismatches.

Purpose of the Study:

• To propose a novel multitask evolutionary framework for effective MTO.
• To address the challenge of differing solution distributions across tasks in MTO.
• To enhance knowledge aggregation and online learning for improved MTO performance.

Main Methods:

• A domain adaptation-based mapping strategy is employed to minimize differences across solution domains.
• The framework facilitates effective information interactions by identifying shared genetic traits.
• A population division strategy and targeted learning approach are used to optimize individual learning from other tasks.

Main Results:

• The proposed framework effectively reduces distribution differences between task solutions.
• Improved information interactions and knowledge aggregation were observed.
• Empirical studies on MTO benchmarks demonstrate superior performance compared to state-of-the-art methods.

Conclusions:

• The novel multitask evolutionary framework offers a robust solution for MTO problems with diverse task distributions.
• Domain adaptation and strategic population management are key to successful intertask learning.
• The proposed paradigm significantly advances the state of the art in multitask optimization.