A Preliminary Study of Knowledge Transfer in Multi-Classification Using Gene Expression Programming.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Gene Expression Programming (GEP), a variant of Genetic Programming (GP), is widely used for automatic program generation and classification.
• Standard GEP approaches multi-classification by decomposing it into multiple binary classification tasks, often ignoring inter-class relationships.
• This decomposition can lead to output conflicts and degraded performance in multi-class scenarios.

Purpose of the Study:

• To address the output conflict issue in GEP-based multi-classification.
• To enhance the performance of GEP multi-classifiers by leveraging evolutionary multitasking optimization.
• To investigate the effectiveness of knowledge transfer strategies among separate binary GEP classifiers.

Main Methods:

• Implementation of an evolutionary multitasking optimization paradigm within an existing GEP multi-classification framework.
• Integration of several knowledge transfer strategies to facilitate interaction between populations of separate binary GEP classifiers.
• Evaluation of the proposed approach on 10 high-dimensional datasets.

Main Results:

• The proposed method effectively alleviates output conflicts among individual binary GEP classifiers.
• Knowledge transfer among separate binary classifiers demonstrably enhances multi-classification performance.
• Improved performance was achieved within the same computational budget compared to traditional methods.

Conclusions:

• Evolutionary multitasking and knowledge transfer are effective strategies for improving GEP-based multi-classification.
• Addressing inter-class relationships through knowledge sharing mitigates performance degradation.
• The findings suggest a promising direction for developing more robust GEP multi-classification systems.