2013 | 21 | 2(82) | 75-98
Article title

Maszyny a metoda naukowa

Title variants
Languages of publication
The paper explores possible influences that recent developments in the field of a branch of AI called Automated Discovery Systems might have upon some aspects of the old debate between Francis Bacon’s inductivism and Karl Popper’s falsificationism. Francis Bacon advocates mechanical induction as the legitimate, infallible method of science, and Karl Popper proposes his famous falsificationist view, according to which science proceeds by subsequent conjectures and refutations, and the question about where scientific hypotheses come from neither needs, nor is capable of, logical analysis. The traditional method of discussing such methodological debates relies on the analysis of various historical examples of discoveries in order to see how well the two models of scientific method account for them. A British philosopher of science and historian of mathematics, Donald Gillies, argues, after some analysis of historical cases of discovery, that Baconian induction had been used in science very rarely, or not at all, although the situation has changed with the advent of a branch of AI called machine learning systems. I think that Gillies’s line of argument can be generalized. Thanks to Machine Discovery Systems, philosophers of science have at their disposal a new tool for empirically testing their philosophical hypotheses: the measure of success or failure of philosophical conceptions about science is how well computer discovery systems that incorporate them in their heuristic principles perform in making scientific discoveries. Accordingly, in the paper I make an attempt to answer which of the two philosophical conceptions of scientific method is better vindicated in view of the successes and failures of systems developed within three major research programs in the field: machine learning systems in the Turing tradition, normative theory of scientific discovery formulated by Herbert Simon’s group, and the program called HHNT, proposed by J. Holland, K. Holyoak, R. Nisbett, and P. Thagard. I agree with Donald Gillies that Baconian induction incorporating, to some extent, Popper’s ideas of falsifying and rejecting hypotheses really did become part of scientific method. It is used in rules of inference of machine learning systems producing hitherto unknown scientific laws. These laws, however, are very low-level statistical rules describing specific phenomena and it can be questioned whether they deserve the name of full-fledged scientific laws. In Simon’s tradition, mechanical induction, contrary to the critics of that program, including Gillies himself, is used by systems that discover the hidden structure of matter or formulate complicated models of processes. Those systems use theory-poor methods which are quite distinct from high-level theoretical methods that were actually used by human discoverers in the field. Moreover, those systems generate models of the hidden micro-structure in purely mechanical manner, but then, in using heuristics to cut down on complexity by reducing the search as early as possible, they reject inadequate models in accord with Popper’s methodology. The HHNT program is still under development and, as yet, it lacks practical success in terms of working systems making quantitative and not only qualitative discoveries, but it aims at cognitive, conceptual analysis and computer implementations, using inductive methods, of extremely complicated processes involved in autonomous reasoning of a cognitive system making scientific discovery, also that of theoretical character.
Physical description
Document Type
Publication order reference
YADDA identifier
JavaScript is turned off in your web browser. Turn it on to take full advantage of this site, then refresh the page.