A Two-Level Transfer Learning Algorithm for Evolutionary Multitasking.

Area of Science:

• Computational intelligence
• Optimization algorithms
• Machine learning

Background:

• Multitasking optimization (MTO) addresses multiple tasks simultaneously, leveraging task similarities for improved performance.
• Existing methods like the multifactorial evolutionary algorithm (MFEA) use basic transfer learning, leading to slow convergence.
• There is a need for more effective transfer learning strategies in MTO.

Purpose of the Study:

• To propose a novel two-level transfer learning (TLTL) algorithm for multitasking optimization (MTO).
• To enhance the efficiency and effectiveness of MTO by improving convergence speed and global search.
• To address the limitations of random inter-task transfer learning in existing MTO algorithms.

Main Methods:

• Developed a two-level transfer learning (TLTL) algorithm for MTO.
• Upper-level: Implements inter-task transfer learning using chromosome crossover and elite individual learning.
• Lower-level: Incorporates intra-task transfer learning via decision variable information transfer for across-dimension optimization.

Main Results:

• The proposed TLTL algorithm demonstrates an outstanding global search ability.
• Experimental studies confirm a significantly fast convergence rate compared to existing methods.
• The algorithm effectively utilizes correlations and similarities among component tasks.

Conclusions:

• The TLTL algorithm offers a superior approach to MTO compared to traditional methods.
• This method enhances optimization efficiency and effectiveness by exploiting task relatedness.
• The findings suggest a promising direction for advancing multitasking optimization research.