Skip to content

GP Subtree Transfer Learning Bottleneck

Home / Open Questions / GP Subtree Transfer Learning Bottleneck

Background: Genetic programming solutions are typically evolved from scratch for specific datasets, which incurs significant computational overhead in data-constrained or large-scale settings. Increasing search efficiency via structural reuse remains a challenge.

Question / Future Work: Investigate methods for transferring or composing successful subtree building blocks across related time series classification tasks. This requires developing mechanisms to identify, regularize, and adapt cross-dataset feature modules to improve convergence speed and generalizability.

Why It Matters: Addressing the computational cost and cold-start problem of GP models through transfer learning is a key step toward making them competitive with pre-trained foundation models.

Metadata & Links