Morphological development guiding evolution (Baldwin effect and self-scaffolding)

Child (1915) in Individuality in Organisms held that developmental gradients appear as expressions of the metabolic axial gradients. He wrote that they usually become more distinct as morphological development proceeds.¹ He read the so-called law of antero-posterior development as a partial recognition of this. Organs first become visible at the anterior end, and morphogenesis then proceeds toward the posterior in an orderly way.¹ Thompson, in On Growth and Form, framed form through rates of growth in different directions. When the ratio between those rates alters, the result is morphological development, a steady and persistent change of form.²

Kriegman, Cheney and Bongard (2018) counted only four reports of simulated robots whose bodies changed while they behaved. In the first two, it was unclear whether the change helped the evolution of behaviour.³ Bongard then showed a form of self-scaffolding that smoothed the fitness surface and raised evolvability.³ The same change exposed evolution to more sensor-motor contingencies and raised robustness in novel environments.³ Later work showed development sweeping over a series of body plans in one agent, with heterochronic mutations canalizing the most promising plan in more static descendants.³

In the model of Kriegman and colleagues, more morphological development raises the chance of capturing an advantageous static phenotype, which can then be canalized once found.⁴ They add a limit. Maximizing development does not reach the global optimum, because it only lengthens the line on which development unfolds in phenotypic hyperspace.⁴ They also note that plasticity increased before genetic assimilation without any environmental change. Models of the Baldwin effect usually assume that plasticity itself does not evolve.⁴

Levin (2022) drew an analogy to the Baldwin effect. Competent cells hide the negative consequences of a mutation, so a lineage can find out what useful things the mutation does.⁵ This gives time for later mutations to hardwire the compensatory changes. Shreesha and Levin (2023) cite Hinton and Nowlan, who found that adding learning to evolutionary simulations gave results consistent with the Baldwin effect.⁶ Their own first simulations of cellular competency did not show it. They suggested this might be because the model set no cost on competency.⁶ They advise re-investigating the question once real costs can be measured.⁷ In planaria, the same wild-type cells can hold different bioelectric representations of head number, Levin (2023) notes.⁸