Learning Task Relationships in Evolutionary Multitasking for Multiobjective Continuous Optimization.

Area of Science:

• Artificial Intelligence
• Optimization Algorithms
• Machine Learning

Background:

• Multiobjective multifactorial optimization (MO-MFO) addresses solving multiple distinct multiobjective optimization problems concurrently.
• Existing methods face challenges in effectively leveraging relationships between diverse tasks.

Purpose of the Study:

• To propose a novel evolutionary multitasking algorithm with learning task relationships (LTR) for MO-MFO.
• To enhance information transfer across heterogeneous decision spaces in MO-MFO.

Main Methods:

• Modeling task decision spaces as manifolds and jointly representing them as a single joint manifold.
• Projecting the joint manifold to a latent space using generalized eigenvalue decomposition.
• Utilizing a joint mapping matrix derived from the latent space projection for cross-task information transfer.

Main Results:

• The proposed LTR algorithm demonstrated superior performance compared to state-of-the-art solvers on various MO-MFO test problems.
• The algorithm effectively handled complex MO-MFO scenarios with heterogeneous decision spaces.

Conclusions:

• The LTR approach offers a significant advancement in MO-MFO by effectively learning and utilizing task interdependencies.
• This method provides a robust solution for tackling complex, multi-task optimization challenges.