Dynamic primitive granularity control
Braden N. Tisdale, Aaron Scott Pope, Daniel R. Tauritz · 2020
Dynamic primitive granularity control (DPGC) is a promising avenue for improving the performance of genetic programming (GP). However, it remains almost entirely unexplored. Further, it may pose many unique challenges in its design and implementation that traditional GP implementations do not. This paper presents an implementation of DPGC in order to determine what aspects of conventional GP design and implementation require special consideration. There are some common techniques used in GP that have been found here to negatively impact DPGC's ability to improve performance. Parsimony pressure appears to disproportionately penalize low-level primitives, and a mixed-granularity population suffers from heavy biases towards particular granularity levels, seemingly to the detriment of evolution. This paper provides hypotheses as to why these conventional techniques harm DPGC implementations, as well as several potential alternatives for use in the future that may remedy these detrimental effects.