Survival trees are a statistical modeling technique used to predict the time until an event occurs. They are widely valued for their interpretability, as they allow practitioners to understand how different variables influence outcomes. However, traditional survival trees struggle to capture nonlinear relationships and rely on greedy splitting strategies, which limit their performance in complex settings.

This thesis proposes a novel approach that addresses these limitations by generating globally optimized survival trees using Genetic Programming Gene-pool Optimal Mixing Evolutionary Algorithm (GP-GOMEA). By integrating a state-of-the-art evolutionary algorithm into the tree construction process, the resulting survival trees optimize both the structure and the decision nodes at a global level.

The method was evaluated on a synthetic dataset designed to require nonlinear decision boundaries—the XOR problem. Our approach consistently outperformed traditional survival trees, achieving optimal or near-optimal results in the noise-free setting. Moreover, the results show that GP-GOMEA survival trees can maintain a high performance even with a smaller population size and limited data, demonstrating the method’s suitability for problems involving nonlinear interactions.

These findings suggest that GP-GOMEA survival tree is a promising direction for advancing survival tree methodology. Future work should include evaluating the method on real-world survival datasets and further tuning key hyperparameters, such as the number of decision nodes.