Approximating distributions of closed-form mathematical expressions for symbolic regression

Identifier:  TFG:9494

Handle: https://hdl.handle.net/20.500.11797/TFG9494

Authors:  Nabaza Ruibal, Leonel Fernando

Abstract:
This thesis presents a new approach to improve symbolic regression model sampling by introducing the probability tree, a structure that samples expressions based on node-specific distributions. Using a Bayesian Machine Scientist (BMS), the tree is trained using models from both the BMS and itself by minimizing the Kullback–Leibler divergence between description lengths and assigned probabilities. Results show that the tree effectively learns to guide expression sampling, enabling a more efficient model space exploration. This suggests that probability trees are a promising alternative or complement to traditional symbolic regression techniques.