Download Full Text (1.4 MB)


In evolutionary computation, programs are developed using evolution's basic principles, such as selection, mutation, and recombination, to iteratively improve problem solutions towards optimal outcomes in a reasonable amount of time. To save time and be more efficient, we are currently exploring a modified version of phylogeny-informed fitness estimation. The original version evaluates each individual program on a subset of the training cases and estimates the performance everywhere else according to its parent's performance. Our approach involves comprehensive evaluation of promising programs across all training cases, increasing computational investment where the sub-sampled results indicated potential gains. This method led to our modified algorithms finding solutions in fewer generations, but at the cost of increased computation time. One question is how to determine whether a solution is promising enough to warrant this additional evaluation. To address this, we used a threshold-k, requiring that a child should be better than its parent in at least k training cases. Analysis of 30 trials on a simple test problem showed threshold-3 enhanced time efficiency, while threshold-1 minimized the number of generations needed for success. A further 100-run analysis with much lower generation limits revealed that threshold-1 secured the highest success rates. To sum up, we did a lot of interesting experiments using a modified version of phylogeny-informed fitness estimation, including comparisons between modified version and original one, 30 runs and 100 runs, complex regression problem and fuel-cost problem.

Publication Date



Evolutionary computation; Genetic programming; Parent selection; Phylogeny; Lexicase selection


Numerical Analysis and Scientific Computing


Author: Chenfei Peng
Project Advisor: Nicholas McPhee

Enhancing Evolutionary Computation: Optimizing Phylogeny-Informed Fitness Estimation Through Strategic Modifications