Scaling Multiobjective Evolution to Large Data With Minions: A Bayes-Informed Multitask Approach.

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Machine Learning

Background:

• Digitalization increases data volume and variety, challenging data-driven optimization.
• Scalable multiobjective evolution is crucial for large-scale data problems.

Purpose of the Study:

• To develop an efficient evolutionary multitasking framework for large-instance data optimization.
• To design a computational resource allocation strategy to leverage auxiliary tasks effectively.
• To mitigate negative transfer between tasks.

Main Methods:

• Utilizing subsampled small-data tasks as auxiliary source tasks (minions).
• Implementing an evolutionary multitasking framework with a novel resource allocation strategy.
• Defining and approximating an intertask empirical correlation measure via Bayes' rule for online resource allocation.

Main Results:

• The proposed algorithm demonstrated significant speedups, up to 73%, compared to existing methods.
• Effective utilization of auxiliary tasks was achieved while preventing negative transfer.
• The approach was validated on benchmark problems and deep neural network hyperparameter tuning.

Conclusions:

• The proposed evolutionary multitasking framework efficiently optimizes large datasets using auxiliary tasks.
• The novel resource allocation strategy ensures effective learning from source tasks.
• This method offers a substantial performance improvement for real-world multiobjective optimization problems.