Learning to Transfer for Evolutionary Multitasking.

Area of Science:

• Artificial Intelligence
• Optimization Algorithms
• Machine Learning

Background:

• Evolutionary multitasking (EMT) uses knowledge transfer (KT) for solving multitask optimization problems (MTOPs).
• Current implicit EMT methods struggle with adaptability due to limited operator use and insufficient state utilization for KT.
• This leads to underutilizing the potential of implicit KT for diverse MTOPs.

Purpose of the Study:

• To propose a novel learning-to-transfer (L2T) framework for automatically discovering efficient KT policies in EMT.
• To conceptualize KT as a learning agent's strategic decisions within the EMT process.
• To enhance the adaptability and performance of implicit EMT for a wide range of MTOPs.

Main Methods:

• Developed an L2T framework with action/state/reward formulations and an interaction environment.
• Employed an actor-critic network trained via proximal policy optimization for the learning agent.
• Integrated the learned agent with various evolutionary algorithms to handle unseen MTOPs.

Main Results:

• The L2T framework demonstrated significant improvements in adaptability and performance.
• Empirical studies on synthetic and real-world MTOPs validated the framework's effectiveness.
• The approach successfully addressed diverse intertask relationships, function classes, and task distributions.

Conclusions:

• The proposed L2T framework enhances implicit EMT by enabling automated discovery of efficient KT policies.
• This approach overcomes limitations in adaptability and operator utilization of existing methods.
• The L2T framework offers a promising direction for improving the performance of evolutionary algorithms on unseen MTOPs.