Goal assembly as a formalism of evolvable design

Abstract

In the age of artificial intelligence and biotechnology, a unified understanding of technology and biology is critically needed but still lacking. A cornerstone of such unification is evolvable design. I present a formalism, called goal assembly, that unifies three key characteristics of evolvable, biology-like design: (i) hierarchical integration of small competencies into larger competencies; (ii) highly non-uniform (power-law or log-normal distributed) representation of phenotypes by genotypes; and (iii) evolution of hierarchical modularity. It does so by making the hierarchy of physical goal states corresponding to substructures explicit, focusing on their composition across scales. In particular, higher-level goal variables approximate achievable joint goal states of interacting lower-level goal variables. Mechanisms of evolvability include goal state gradient backpropagation across scales, hierarchical decision making among alternative goal states, and structural recombination of modules. In learning theory terms, goal assembly proposes architectural inductive biases of emergent engineering systems operating at a complexity regime where goals arise as necessary abstractions. Beyond asking what goals are and how they are achieved, goal assembly suggests we can almost independently talk about their compositionality.