Search This Blog

Sunday, February 21, 2016

Below the fold I reproduce the text of a recent comment by me on the issue of prediction vs retrodiction as validation of a scientific theory. It is a fundamental point, on which there is much confusion, so it bears preservation.

FrankShann @3, in logical terms, a set of propositions, x, predicts another set of propositions, y, if and only if y can be logically deduced from x.  This is the fundamental relationship that underlies all explanation.  Of course, sometimes we are not able to predict events from a set of propositions, but only the statistical distribution in which the event lies, or in other words, the probability of its occurence.  Being human, we will often claim that something "explains" something else, when it only explains why the event is highly probable - but that does not alter the fact that fundamentally, explanation is logical deduction.
The sole difference between prediction and retrodiction is that the former is explanation before the event, and the later is explanation after the event.  Logically, this is irrelevant to how impressive the explanation is.  One explanation is superior to the other based on simplicity (ie, the number of entities and relationships invoked), the preciseness of the conclusion of the successful deduction, and a priori probability of the premises.  Nothing else, including the time it was made, enters into the fact.  We are not less impressed by Newton's deduction of Galilean kinematics from his laws of motion, nor of Keppler's laws of planetary motion from his laws of motion plus the law of universal gravitation because they were after the event - and nor should we be.
The reason we are suspicious of retrodiction is the suspicion that they are ad hoc, ie, that they relly on premises added after the event to make the prediction fit, and at the cost of the simplicity of the premises used.  However, the inclussion of ad hoc premises can be tested for either before or after the event.  Therefore, provided we exclude ad hoc premises, prediction is no better in a scientific theory than retrodiction.  Indeed, that is necessarilly the case in science.  Otherwise we would need to preffer a theory that made correct predictions into the future but entirely failed to retrodict past observations over a theory that both predicted and retrodicted past and future observations with a very high degree of accuracy but occasional failures.  Indeed, as we cannot know in advance future success, science is built on the principle that successful retrodiction in the best guide to successful prediction.
Given the above, your suspicions of CMIP5 models is based on an assumption that the change between them and earlier models is from the addition of ad hoc premises.  That is in fact contrary to the case.  The earliest climate models, due to lacking perfect resolution, needed ad hoc adjustments to close the energy budget.  They needed ad hoc values for the rate of heat absorption by the ocean because they did not model the ocean.  The very earliest models required ad hoc assumptions about the ratio of increase of different GHG because they did not have the capacity to model all GHG.  As computer power has been improved, these ad hoc assumptions have been progressively removed.  In terms of the elegance of prediction, CMIP5 models are vastly preferrable to the older models - but that is the crucial criteria.
If we prefer the predictions of Hansen (88) as a test of the validity of climate science - we are being unscientific.  The model used in Hansen (88) did not include aerosols, did not include all GHGs, used a swamp ocean, did not include a stratosphere, and was not able to be run enough to generate an ensemble of predictions (a necessary feature for generating the probabilistic predictions of climate).  In short, it was a massively ad hoc model, especially when compared to its modern incarnation.  Therefore, if we are interested in science rather than rhetoric, the successful retrodiction by CMIP 5 models should impress us more than successful (or unsuccessful) predictions of Hansen (88).
Nor is the development from more use of ad hoc premises to less either unusual or a problem in science.  In fact it is typical.  Newton started predicting the motion of planets using the ad hoc premise that planets were point masses.  Later that was improved upon by the ad hoc premise that planets were empty shells with all their mass distributed evenly at their surface.  Only as computational power and mathematical techniques have improved has it become possible to model planets as genuine 3-D objects with variable mass concentrations in Newton's theory.  This was not a basis of rational criticism of Newton's theory, and nor is the primitive nature of the model used in Hansen (88) a valid criticism of climate science.  But just as we would not prefer continuing to use point masses in prediction in gravitation, nor should we preffer the predictions of Hansen (88) over the retrodictions of CMIP5.

No comments:

Post a Comment