Geometric Semantic Genetic Programming Is Overkill
[ 1 ] Instytut Informatyki, Wydział Informatyki, Politechnika Poznańska | [ P ] employee
2016
paper
english
- automatic program induction
- geometric semantic genetic programming
- solution space
EN Recently, a new notion of Geometric Semantic Genetic Programming emerged in the field of automatic program induction from examples. Given that the induction problem is stated by means of function learning and a fitness function is a metric, GSGP uses geometry of solution space to search for the optimal program. We demonstrate that a program constructed by GSGP is indeed a linear combination of random parts. We also show that this type of program can be constructed in a predetermined time by much simpler algorithm and with guarantee of solving the induction problem optimally. We experimentally compare the proposed algorithm to GSGP on a set of symbolic regression, Boolean function synthesis and classifier induction problems. The proposed algorithm is superior to GSGP in terms of training-set fitness, size of produced programs and computational cost, and generalizes on test-set similarly to GSGP.
246 - 260