Source: Weber, Marcel. "The Crux of Crucial Experiments: Duhem's Problems and Inference to the Best Explanation." Brit. J. Phil. Sci. 60, 19-49
Pierre Duhem is famous for his severe philosophical criticism of severe testing, or crucial experimentation, in science. A crucial experiment would follow a form of eliminative induction as follows.
Second, while we can purport to parse the full field of contenders in logic or mathematics, we cannot do so in physics or economics. According to Duhem, “Shall we ever dare to assert that no other hypothesis is imaginable?”
The answer, of course, is no. But Duhem’s dire turn of phrase should not deter us from saying, "Big deal," and forging on with science uncowed. We might accept a pragmatic constraint on science as necessitating or justifying a means of distinguishing between theories, because the alternative, anarchy, is unacceptable.
Said differently, Duhem’s epistemic standards are too high: imaginability is not a useful criterion with which to judge science.
The conventional answer as I have phrased it is pick the best hypothesis, call it theory, and move on. This is called inference to the best explanation and it is the line taken by Marcel Weber.
The conventional rejoinder comes as a finer distillation of the Duhem critique. This is Bas van Fraassen’s “bad lot”-argument: if all the choices are crappy, even the least crappy is still crappy. Well, crap.
Or not. According to Weber, experiments do not refute hypotheses for contradiction with evidence – data produced by the experiments themselves – they select for maximal consonance with it. Weber calls this inductive experimental inference to the best explanation (IBE).
To explain how this works, Weber exhaustively describes controversy around the Meselson-Stahl experiment of the Watson-Crick semi-conservative DNA reproduction hypothesis. He also, perhaps unnecessarily, involves mechanism talk.
Weber defines experiments as mechanisms: experimental mechanisms. These are human-designed, generalizing-producing, contingent causal processes. Hypotheses are also mechanisms: theoretical mechanisms (although Weber does not say so). Then there are physiological mechanisms. These are embedded in experimental mechanisms. They're real and natural causal processes.
Finally, theoretical mechanisms are explanations when they map onto one another and when their inferences match the data produced by experimental mechanisms. The criterion is efficiency: if the whole theoretical mechanism is represented in the whole experimental getup, and these match the experimental data, then we have achieved explanatory sufficiency.
Auxiliary assumptions, in this view, “hitch a free inferential ride on the experimental mechanism, powered by the latter’s explanatory force” (39).
Weber makes the point that we don't care if this line-up is fallible because we are humans. IBE gives adequacy requirements not infallibility.
Now, to the problem of untested auxiliaries. Mechanisms are hierarchical or have levels. Experimental mechanisms (which I have called “the getup”) have sub-mechanisms or constituent causal processes, including physiological mechanisms. Experimentalists tinker with these the same way theorists tinker with theoretical assumptions. Models are models.
However, these sub-mechanisms should be transparent and testable in the experimental getup, and even when there are multiple potential mechanisms behind regularities, these controversies can be resolved by IBE. That is, "IBE turtles all the way down."
According to Weber, although Duhem’s problem applies, it is solved by iterated application of the IBE because explanatory sufficiency in the IBE framework does not require the vanquishing of all auxiliary hypotheses, or, in this case, mechanisms. Nobody disagrees. So this is a straw man argument.
Nonetheless, Weber avers that there will be a winner. One out of the pack of competing hypotheses will more or less hit the nail on the head. The others will require fiddling in the form of additional (ad hoc) assumptions, and the difference between these types of hypotheses will be transparent.
All of this is predicated on the cogency of the mechanism talk. Mechanisms provide constraints. This argument explicates other solutions to the bad lot problem. For instance, the simpler solution to the bad lot complaint was that scientists can (mostly) exhaust the space of possible (i.e. plausible) hypotheses. Although Weber rejects this argument by Roush (2005) it is in fact his argument, just dressed up in mechanisms.
The original argument relies on the intuition that hypotheses must meet sufficiency constraints. Weber updates these as mechanistic constraints and describes them as “very stringent” (44). These are identified by Craver (2007): componency, spatial, temporal, and active constraints, to which Weber adds functional constraints. I have no idea what these mean, but they should arise, innocuously, from “background knowledge” (45).
As an extension, just to make sure that he has discussed everything in the philosophy of science at the same time, Weber shows how this account is consistent with Bayesian theory. I skip that and go directly to my qualms.
Weber’s account does satisfy my so-called pragmatic condition. But it is only slightly less inductively eliminative than Duhem’s. Sufficiency is useful, but only when it pertains to last man standing hypotheses. If sufficiency pertains to multiple hypotheses, and there is no clearly ad hoc adjustment – or even if there is – their contest is indeterminate.
Further, the mechanistic inclusion is gratuitous. Calling experiments experimental mechanisms with embedded physiological mechanisms doesn’t change the content or relation of experiments and the world nor our access to it. Mechanisms are epistemic heuristics, not ontological furniture.
Finally, and related to the last point, Weber rolls roughshod over what I consider to be adequate, albeit old and dry, solutions to Duhem’s problems: recognizing that Duhem has over-exaggerated epistemic standards and the feasibility of exhausting plausible hypotheses of data.
Weber’s account of the Meselson-Stahl and Watson-Crick debates is excellent and seems to back up all his points, but my critique of the conceptual apparatus stains this praise.
Pierre Duhem is famous for his severe philosophical criticism of severe testing, or crucial experimentation, in science. A crucial experiment would follow a form of eliminative induction as follows.
Starting with alternate hypotheses of an evidenced phenomena, an experiment could negate one or all but one by showing contradiction with the evidence. This constitutes modus tollens or denying the consequent of the form “If h1, then e. Not e. Therefore, not h1.” If e follows from h2, then by disjunctive syllogism we can conclude the veracity of h2, or the last hypothesis standing.
Such a train of inference from experimentation is admissible according to logic, but it faces two difficulties in practice.
First, hypotheses are not singular postulates but conjunctive bundles. They rely on auxiliary hypotheses which rely on auxiliary hypotheses which rely on auxiliary hypotheses and so on... This is called the problem of untested auxiliaries.
Such a train of inference from experimentation is admissible according to logic, but it faces two difficulties in practice.
First, hypotheses are not singular postulates but conjunctive bundles. They rely on auxiliary hypotheses which rely on auxiliary hypotheses which rely on auxiliary hypotheses and so on... This is called the problem of untested auxiliaries.
Eliminative induction fails because humans cannot pinpoint error; they cannot grab the tip of the devil’s tail.
Second, while we can purport to parse the full field of contenders in logic or mathematics, we cannot do so in physics or economics. According to Duhem, “Shall we ever dare to assert that no other hypothesis is imaginable?”
The answer, of course, is no. But Duhem’s dire turn of phrase should not deter us from saying, "Big deal," and forging on with science uncowed. We might accept a pragmatic constraint on science as necessitating or justifying a means of distinguishing between theories, because the alternative, anarchy, is unacceptable.
Said differently, Duhem’s epistemic standards are too high: imaginability is not a useful criterion with which to judge science.
The conventional answer as I have phrased it is pick the best hypothesis, call it theory, and move on. This is called inference to the best explanation and it is the line taken by Marcel Weber.
The conventional rejoinder comes as a finer distillation of the Duhem critique. This is Bas van Fraassen’s “bad lot”-argument: if all the choices are crappy, even the least crappy is still crappy. Well, crap.
Or not. According to Weber, experiments do not refute hypotheses for contradiction with evidence – data produced by the experiments themselves – they select for maximal consonance with it. Weber calls this inductive experimental inference to the best explanation (IBE).
To explain how this works, Weber exhaustively describes controversy around the Meselson-Stahl experiment of the Watson-Crick semi-conservative DNA reproduction hypothesis. He also, perhaps unnecessarily, involves mechanism talk.
Weber defines experiments as mechanisms: experimental mechanisms. These are human-designed, generalizing-producing, contingent causal processes. Hypotheses are also mechanisms: theoretical mechanisms (although Weber does not say so). Then there are physiological mechanisms. These are embedded in experimental mechanisms. They're real and natural causal processes.
Finally, theoretical mechanisms are explanations when they map onto one another and when their inferences match the data produced by experimental mechanisms. The criterion is efficiency: if the whole theoretical mechanism is represented in the whole experimental getup, and these match the experimental data, then we have achieved explanatory sufficiency.
Auxiliary assumptions, in this view, “hitch a free inferential ride on the experimental mechanism, powered by the latter’s explanatory force” (39).
Weber makes the point that we don't care if this line-up is fallible because we are humans. IBE gives adequacy requirements not infallibility.
Now, to the problem of untested auxiliaries. Mechanisms are hierarchical or have levels. Experimental mechanisms (which I have called “the getup”) have sub-mechanisms or constituent causal processes, including physiological mechanisms. Experimentalists tinker with these the same way theorists tinker with theoretical assumptions. Models are models.
However, these sub-mechanisms should be transparent and testable in the experimental getup, and even when there are multiple potential mechanisms behind regularities, these controversies can be resolved by IBE. That is, "IBE turtles all the way down."
According to Weber, although Duhem’s problem applies, it is solved by iterated application of the IBE because explanatory sufficiency in the IBE framework does not require the vanquishing of all auxiliary hypotheses, or, in this case, mechanisms. Nobody disagrees. So this is a straw man argument.
Nonetheless, Weber avers that there will be a winner. One out of the pack of competing hypotheses will more or less hit the nail on the head. The others will require fiddling in the form of additional (ad hoc) assumptions, and the difference between these types of hypotheses will be transparent.
All of this is predicated on the cogency of the mechanism talk. Mechanisms provide constraints. This argument explicates other solutions to the bad lot problem. For instance, the simpler solution to the bad lot complaint was that scientists can (mostly) exhaust the space of possible (i.e. plausible) hypotheses. Although Weber rejects this argument by Roush (2005) it is in fact his argument, just dressed up in mechanisms.
The original argument relies on the intuition that hypotheses must meet sufficiency constraints. Weber updates these as mechanistic constraints and describes them as “very stringent” (44). These are identified by Craver (2007): componency, spatial, temporal, and active constraints, to which Weber adds functional constraints. I have no idea what these mean, but they should arise, innocuously, from “background knowledge” (45).
As an extension, just to make sure that he has discussed everything in the philosophy of science at the same time, Weber shows how this account is consistent with Bayesian theory. I skip that and go directly to my qualms.
Weber’s account does satisfy my so-called pragmatic condition. But it is only slightly less inductively eliminative than Duhem’s. Sufficiency is useful, but only when it pertains to last man standing hypotheses. If sufficiency pertains to multiple hypotheses, and there is no clearly ad hoc adjustment – or even if there is – their contest is indeterminate.
Further, the mechanistic inclusion is gratuitous. Calling experiments experimental mechanisms with embedded physiological mechanisms doesn’t change the content or relation of experiments and the world nor our access to it. Mechanisms are epistemic heuristics, not ontological furniture.
Finally, and related to the last point, Weber rolls roughshod over what I consider to be adequate, albeit old and dry, solutions to Duhem’s problems: recognizing that Duhem has over-exaggerated epistemic standards and the feasibility of exhausting plausible hypotheses of data.
Weber’s account of the Meselson-Stahl and Watson-Crick debates is excellent and seems to back up all his points, but my critique of the conceptual apparatus stains this praise.
No comments:
Post a Comment