Universal Darwinism as Bayesian inference

Rádl (1930) gives an early account of how Darwin treated difference. He stressed that to Darwin the differences between species, varieties and individuals were quantitative only. Rádl wrote that Darwin was content merely to prove that such transitions exist.¹ Rádl also noted that Darwin knew Linnaeus believed in the constancy of transitional forms and had concluded that species are immutable. According to Rádl, Darwin did not seem to think it necessary to ask how Linnaeus came to hold so different a view.¹

McGregor, Vasas, Husbands and Fernando (2012) applied Bayesian reasoning to evolved chemical networks. They did not use it as an experimenter refining their own beliefs. They asked what beliefs a system should hold about its environment if it were itself a rational observer. They found that a Bayesian analysis gave insight into network function. There was no explicit pressure on the networks to perform Bayesian reasoning.² High fitness during evolution did require the networks to integrate information over time. The authors took iterated Bayesian inference to be the formal ideal of integrating cumulative evidence, and used this as their motivation for the Bayesian reading.²

Campbell (2016) argues from the Price equation. It is traditionally understood as the mathematics of biological evolutionary change. Campbell notes that it may be derived from a form of Bayes' theorem, citing Gardner (2008), Shalizi (2009) and Frank (2012b). On that basis it describes Bayesian inference in a very general form.⁵ Campbell states that it splits change into two parts. Selection tends to decrease model variation. Transmission, or copying of the model, serves to increase it.⁶ In biology transmission change is often genetic mutation. Campbell expects an analogous variation-raising mechanism in non-biological processes such as scientific models. He concedes that how new scientific models arise is less clear, and that it may seem to have little in common with random mutation.⁶

Other authors use Bayesian inference to describe agents and living systems. Friston and colleagues (2013) argue that agents must perform active Bayesian inference. Exact inference is generally intractable and cannot be realized biophysically, so the problem is how agents perform approximate inference. They contrast this with classical formulations in which agents optimize expected value.³ Friston (2013) argues that if internal states are read as encoding Bayesian beliefs, the dynamics of internal and active states can be described as gradient descent on a variational free energy.⁴ Ramstead and colleagues (2019) place the free energy principle within an evolutionary systems theory.⁷ Pio-Lopez and colleagues (2022) describe the bound that underlies Bayesian inference in this principle.⁸