Two Senses of Robustness and Hacking’s Scientific Realism

As already mentioned, it is common to distinguish different senses of the term “robustness”. For example, drawing on Wimsatt (2007), Calcott (2010) distin­guishes the three following meanings:

(1) Robustness of models.

(2) Robustness as stability or insensitivity of output. In this sense, a phenomenon, mechanism or system may be said to be robust, if it shows a relative stability or insensitivity of output as against variations in parameter values. In this sense, robustness is important both in engineered systems and in living systems because it provides resilience against internal or external perturbations.

(3) Robustness as consilience of results coming from different and independent sources of evidence. These latter are said to be robust if they are supported in a variety of independent ways. It is this sense of robustness that is emphasized by those who espouse the no miracles argument.

I do not wish to discuss meaning (1) in this paper. I shall say something about the last two meanings of robustness, and I shall call them robustness-as-stability and robustness-as-consilience.^[49]

According to the first sense that we shall discuss, robustness is usually defined as “a property that allows a system to maintain its functions against internal and external perturbations.” (Kitano 2004, 826; cf. also Stelling et al. 2004, and Clausing 2004, 25). In this sense, robustness is very important in engineered sys­tems, as it makes them more resistant to events that cannot de facto (and perhaps in principle) be predicted in the early stages of their development: think of a robot that must work in still unexplored portions of nature, for example. This kind of robustness is usually obtained by building redundancy into a mechanism: if one element fails, another element can play its role (cf.

In this same sense, robustness is also a very important property in living systems, as resilience against internal or external perturbations. For example, the genetic code can be described as a robust encoding of amino acids into codons, or we may say that proteins, developmental pathways, metabolic networks, and tumours are robust against, respectively, translation errors, environmental or genetic distur­bances (e.g. “gene knockout experiments”), changes in enzyme efficiency, and various chemotherapies (cf. Wilke 2006, 695; Strand and Oftedal 2009). This meaning of the term is also exemplified by mechanisms by which living beings increase their survival rate, such as the mechanism of anhydrobiosis by which, under extreme dehydratation, tardigrades suspend metabolism almost completely by extensive production of trehalose and become active again upon rehydration (cf. Crowe and Crowe 2000; Singer and Lindquist 1998).

The second meaning of robustness, that is, robustness as consilience or coin­cidence of a variety of different (independent) pieces of evidence, is both a method which is frequently used in everyday life and a venerable concept in the philosophy of science. Apart from some partial anticipation (on this point, cf. especially Wimsatt 2012; Stegenga 2009; Hudson 2014), it is well-known that William Whewell was the first important author to be fully aware of the importance of this concept, which he called “Consilience of Inductions’” (cf. Whewell [1840] 1847, Vol. 2, 65—66). On the contrary, Bridgman (Bridgman 1927, 56—60, who employed the notion—though not the term—of robustness-as-consilience as a criterion for proving the physical reality of theoretical entities), Popper (for the importance of independent tests: 1963[1972] and 1972), and Glymour (for his “logical pincer movement” or “bootstrap testing”: 1974 and 1980) are to be especially mentioned among the most important authors that have been too much neglected in the robustness debate hitherto.

Newton's theory of universal gravitation is one of the best examples. No relation between Kepler's first, second, and third laws concerning the motion of the planets around the sun was present until Newton's theory of universal gravitation explained all of them at once. Moreover, Newton's theory of universal gravitation did not explain only the perturbations of the moon and planets by the sun and by each other, from which it was originally inferred, but also the apparently independent fact of the precession of the equinoxes (cf. Whewell [1840] 1847, Vol. 2: 65—66). But the most discussed case in the literature is perhaps the measurement of Avo- gadro's number, that is, the number of molecules per gram-mole of any gas, which Jean-Baptiste Perrin computed by quite different methods (among which the most important was perhaps that based on Brownian motion) and found to be the same, with a relatively good—but not so good as it is sometimes supposed (cf. Perrin [1916], 87)—approximation to today's accepted value, that is, 6.02,214,179 X 10²³ mol^-1 (cf. Psillos 2011). The fact that all of these measure­ments essentially agreed with the experimental data within the accuracy of the observations would be an unexplainable coincidence, it would be “miraculous”, if each of the measurements referred to an artefact and matter were not composed of molecules and atoms.^[50]

There are two main objections to robustness as used in arguments of this kind for scientific realism. The first is that robustness-as-consilience reasoning does not provide any autonomous epistemic warrant if one lacks minimally reliable obser­vational data or procedures (cf. Hudson 2014, 200). For this reason, the robustness requirement must be integrated by a requirement of “minimal reliability” (Hudson 2014, 18; cf. also Stegenga 2009, 653).

The second objection goes farther. Even if it were true that any single piece of evidence was highly reliable, it still would be very difficult, if not impossible, to ascertain the independence condition for robustness. In fact, it is widely recognized that robustness-as-consilience in science (and the “no-miracle” argument for sci­entific realism) can be credited with epistemic warrant only if the multiple kinds of available and consilient evidence are independent. Wimsatt, for example, recog­nized that “one of the most critical and important problems in the study of robustness analysis”. (Wimsatt [1981] 2012, 83) is “the failure of the different supposedly independent tests, means of detection, models, or derivations to be truly independent” (Wimsatt [1981] 2012, 82).

3