Transfer Learning via Modularity in a Model of Technological Evolution

Most machine learning algorithms ultimately focus on optimizing solutions to a single target function. Coop-tive phenomena that transfer information across distinct environmental niches, however, lie at the heart of the evolution of complex functions in nature and technology, where solutions adapted for one problem are repurposed to solve another, related problem. Boolean functions have become a popular toy model for exploring the dynamics of such processes, and provide insight into new approaches to evolutionary compu-tation. We implemented the a model of combinatorially evolving logic circuits developed by Brian Arthur and Wolfgang Polak in which solutions to boolean functions are encapsulated as modules that can be used to solve other, more complex functions, and began to explore the sort of transfer phenomena its success depends upon. We observed a significant difference in the dynamics of evolution between when the full suite of fitness functions is present, versus when only a few pieces of the selective pressure are active at a time. Future work is needed examine these dynamics in more detail