Sources and References

Primary texts

• Hanson, N.R. Patterns of Discovery: An Inquiry into the Conceptual Foundations of Science. Cambridge University Press, 1958.
• Simon, Herbert A. Models of Discovery and Other Topics in the Methods of Science. D. Reidel, 1977.
• Reichenbach, Hans. Experience and Prediction: An Analysis of the Foundations and the Structure of Knowledge. University of Chicago Press, 1938.
• Peirce, Charles Sanders. “Abduction and Induction.” In Philosophical Writings of Peirce, ed. Justus Buchler. Dover, 1955.
• Kuhn, Thomas S. The Structure of Scientific Revolutions. 3rd ed. University of Chicago Press, 1996.
• Polanyi, Michael. The Tacit Dimension. Doubleday, 1966.
• Nickles, Thomas, ed. Scientific Discovery: Logic and Rationality. D. Reidel, 1980.
• Achinstein, Peter. The Book of Evidence. Oxford University Press, 2001.
• Lipton, Peter. Inference to the Best Explanation. 2nd ed. Routledge, 2004.
• Thagard, Paul. Computational Philosophy of Science. MIT Press, 1988.
• Schickore, Jürgen. “Scientific Discovery.” In Stanford Encyclopedia of Philosophy, ed. Edward N. Zalta, 2018.
• Davey, Kevin. Selected papers on discovery in modern physics (University of Waterloo).

Online resources

• Stanford Encyclopedia of Philosophy: “Scientific Discovery” (Schickore), “Abduction” (Douven), “Incommensurability” (Sankey), “Thomas Kuhn” (Bird).
• Internet Encyclopedia of Philosophy: “Charles Sanders Peirce.”

Chapter 1: Philosophy of Scientific Discovery — An Introduction

1.1 What Is Scientific Discovery?

At its most basic, a scientific discovery (科学发现) is the arrival at a new item of scientific knowledge — a new fact, law, entity, or theoretical framework — that was not previously known or recognized within the scientific community. But this deceptively simple characterization immediately generates philosophical questions that have occupied epistemologists and philosophers of science for well over a century.

Consider the asymmetry between two phases of any scientific episode. In the first phase, a scientist or group generates a hypothesis, proposes a new classification, or encounters an anomalous phenomenon. In the second phase, the community evaluates, tests, and ultimately accepts or rejects the proposal. These phases feel very different in character. The first seems creative, unpredictable, even serendipitous. The second seems rule-governed, intersubjective, and assessable by shared standards. The central question of this course is whether this intuitive asymmetry is philosophically deep or whether it dissolves under careful analysis.

The philosophical study of discovery asks not merely what discoveries happen to occur, but whether there are normative standards — methodological, logical, or computational — by which discovery processes can be evaluated, guided, or even prescribed. This normative ambition distinguishes philosophy of scientific discovery from history of science, though the two disciplines are deeply intertwined.

1.2 Why Discovery Has Been Philosophically Marginalized

For much of the twentieth century, mainstream philosophy of science operated under a professional consensus that scientific discovery was not a proper subject for philosophical analysis. This consensus, associated with the logical empiricist (逻辑经验主义) tradition, rested on an influential conceptual boundary introduced by Hans Reichenbach in 1938: the distinction between the context of discovery (发现的语境) and the context of justification (辩护的语境).

The argument was roughly as follows. Philosophy is concerned with rational reconstruction — with articulating the logical relationships that make beliefs warranted, arguments valid, and theories empirically supported. The process by which a scientist first comes to entertain a hypothesis is a psychological and sociological matter, contingent on biographical accident, cultural influence, and creative intuition. This process has no logical form that philosophy can capture or evaluate. By contrast, the relationship between evidence and hypothesis — whether the evidence warrants belief in the hypothesis — is a logical relationship amenable to rigorous philosophical treatment. Philosophy of science should therefore confine itself to the context of justification and leave the context of discovery to empirical psychologists and historians.

This verdict was so broadly accepted that for several decades the phrase “logic of discovery” was treated as close to an oxymoron. It was only in the 1970s and 1980s — through the work of Hanson, Kuhn, Simon, and a growing community of philosophers and historians of science — that the philosophical study of discovery was rehabilitated as a legitimate and important enterprise.

1.3 Rehabilitating Discovery: The Routes Back

The rehabilitation of scientific discovery as a philosophical topic proceeded along several distinct but complementary lines, and understanding why this rehabilitation happened helps clarify what is philosophically at stake.

First, historians and philosophers of science challenged the sharp separation between discovery and justification by showing that what counts as evidence and what counts as an acceptable hypothesis are not theory-neutral matters. Hanson argued that observation itself is theory-laden, so the distinction between the “merely psychological” context of discovery and the “genuinely logical” context of justification could not be maintained in the clean form the logical empiricists supposed. If the evidential base of science is already theoretically structured, then the process by which theories are formed cannot be cleanly severed from the process by which they are evaluated.

Second, cognitive scientists and AI researchers, most prominently Herbert Simon, argued that discovery processes — far from being irreducibly creative and irrational — could be modeled computationally using heuristic search strategies. If discovery processes are computable, there is no principled reason to exclude them from rational analysis. Simon’s work demonstrated that programs operating under explicit, articulable rules could rediscover major empirical laws, suggesting that the creative mystique surrounding discovery had been philosophically overstated.

Third, philosophers inspired by Peirce pointed to abductive reasoning (溯因推理) — reasoning to the best explanation — as a genuine logical form operative in the generation of hypotheses, not merely in their evaluation. Peirce had called this form retroduction and had argued it was irreducibly distinct from both induction and deduction. If abduction is a genuine logical form, then the generation of hypotheses has logical structure after all, and the positivist exclusion of discovery from logic rests on an impoverished inventory of inferential forms.

Fourth, Kuhn’s structural account of scientific revolutions offered a framework in which discovery was not an atomic event but an extended process involving the recognition of anomaly, the experience of crisis, and a gestalt shift to a new theoretical paradigm. Although Kuhn’s account raised worries about irrationality, it made discovery philosophically visible in a way that galvanized decades of subsequent research.

1.4 Discovery and Creativity

A persistent philosophical puzzle concerns the relationship between discovery and creativity (创造性). The positivist dismissal of discovery rested partly on the assumption that discovery is irreducibly creative and therefore irreducibly irrational — that wherever genuine creativity is operative, there can be no rational reconstruction. But this assumption conflates two questions that should be kept separate.

The first question is descriptive: What cognitive processes actually occur during creative discovery? This is a question for psychology and cognitive science, and it remains actively investigated. Research on creative cognition — on analogical thinking, conceptual blending, and the role of constraints in generating novel ideas — suggests that creative thought is not random but structured, even if its structure differs from deductive reasoning.

The second question is normative: Can creative discovery be evaluated as better or worse by standards that are recognizably rational? This question is independent of the first. Even if the psychological process of creative discovery is complex and not fully conscious, the output of that process — the proposed hypothesis — can be assessed by its coherence with background theory, its explanatory breadth, its simplicity, and its testability. The normative question does not require that the creative process itself be consciously rational, only that its products are subject to rational assessment.

The philosophers who rehabilitated discovery tended to focus on the normative question. Their claim was not that scientists consciously apply logical rules when generating hypotheses, but that the discovery process is amenable to rational evaluation and partial normative reconstruction.

Chapter 2: Discovery vs. Justification — The Reichenbachian Distinction

2.1 Reichenbach’s Original Formulation

In Experience and Prediction (1938), Hans Reichenbach articulated what would become one of the most influential methodological distinctions in twentieth-century philosophy of science. He wrote that epistemology is not concerned with the actual thought processes of the scientist, but with the logical relations that constitute the rational reconstruction of those processes. Accordingly, he distinguished two tasks:

The epistemologist’s task, on Reichenbach’s view, is to provide a rational reconstruction of the justificatory context — to articulate, in formal or quasi-formal terms, the conditions under which evidence confirms a hypothesis, a theory is acceptable, or a belief is rationally warranted. The discovery context is left to empirical psychology.

2.2 The Philosophical Motivation

Reichenbach’s distinction reflects a broader project shared by the logical empiricists: to demarcate science from non-science, to formalize scientific inference, and to protect the objectivity of scientific knowledge from psychologistic contamination. If the validity of scientific conclusions depended on how those conclusions were reached — on the personality of the discoverer, the accidents of the laboratory, or the prevailing cultural climate — then scientific knowledge would seem to inherit the contingency and variability of its origins. By restricting normative epistemology to justificatory relations, Reichenbach secured an objective standard for scientific knowledge that was independent of its causal history.

This motivation was philosophically legitimate and powerful. It echoes Gottlob Frege’s attack on psychologism in logic: the validity of a logical inference does not depend on the psychological processes by which a thinker comes to perform it. The same inferential act — say, modus ponens — is valid whether performed by a genius or a novice, whether in a state of inspiration or routine checking. Reichenbach was extending this Fregean point from logic to empirical epistemology.

2.3 Kuhn’s Challenge to the Distinction

Thomas Kuhn’s The Structure of Scientific Revolutions poses what many regard as the deepest challenge to the Reichenbachian distinction, even though Kuhn does not always frame his argument in those terms. Kuhn’s challenge operates on two fronts.

First, Kuhn argues that the standards of evidential evaluation are themselves paradigm-relative (范式相对的). What counts as confirming evidence, what counts as an acceptable form of explanation, and what counts as a good scientific problem all depend on the paradigm within which scientists are operating. If justificatory standards are not invariant across scientific development, then there is no single “context of justification” that philosophy can analyze independently of the discovery context that produced the paradigm within which those standards operate.

Second, Kuhn’s account of incommensurability (不可通约性) — the thesis that competing paradigms cannot be fully translated into each other’s terms — undermines the idea that evidence can serve as a neutral arbiter between paradigms. If the evidence available in the justificatory context is already organized by paradigmatic assumptions that derive from the discovery context, then the two contexts cannot be cleanly separated.

2.4 Nickles’s Critique: The Constraint Account

Thomas Nickles, in his edited volume Scientific Discovery: Logic and Rationality (1980), pressed what is perhaps the most sustained philosophical criticism of the Reichenbachian picture. Nickles argued that the distinction, even if defensible as an abstract logical point, is insufficient as a philosophy of science because it ignores the constraint structure that governs hypothesis generation.

Nickles’s key insight is that hypothesis generation is not unconstrained. Scientists do not generate arbitrary conjectures and then test them; they generate hypotheses that satisfy background constraints — compatibility with established theory, accord with known experimental results, mathematical tractability, and coherence with the operative problem formulation. These constraints are not merely psychological influences; they have normative force. A hypothesis that violates well-established constraints is not simply less likely to be true; it is not a good scientific hypothesis relative to the current state of the field.

This means that the generation of hypotheses is subject to rational evaluation even before it reaches the stage of formal testing. The constraint-satisfying character of hypothesis generation has logical and normative structure — and that structure belongs to the context of discovery, not merely to the context of justification.

The problem of generation: Even if the context of justification is formally tractable, scientists must first generate hypotheses before evaluating them. If generation is entirely unconstrained — if any hypothesis is as good a starting point as any other — then the method of conjecture-and-test seems radically incomplete as an account of scientific rationality.

The theory-ladenness problem: Hanson argued that the very act of observation — which supplies the evidence for the context of justification — is already shaped by theoretical commitments. If the boundary between observation and theory is porous, then discovery and justification cannot be as cleanly separated as Reichenbach supposed.

The practical interdependence problem: In actual scientific practice, the processes of generation and evaluation are interleaved. Scientists design experiments with hypotheses in mind; the results of evaluation reshape the direction of further discovery. The contexts are not temporally or cognitively distinct phases but mutually constraining aspects of a unified investigative process.

2.5 Is the Distinction Defensible?

In light of these challenges, is the Reichenbachian distinction defensible at all? A balanced assessment suggests that the distinction retains philosophical value when properly construed, but that its proper construal is significantly narrower than its original formulation.

The distinction remains valuable as a logical point: the causal history of a belief is, in the strictest sense, logically irrelevant to its justificatory status. A hypothesis that was discovered by a hallucinogen-induced vision is not thereby false or unconfirmed; its confirmational status depends on its relationship to the evidence, not its psychological origin. This point survives the Kuhnian and Nickles challenges.

What does not survive is the broader methodological verdict: that philosophy of science should restrict itself to justificatory analysis and treat discovery as philosophically uninteresting. The critics have established that the processes of discovery exhibit rational structure, that they are governed by normative constraints, and that philosophical analysis of those constraints is both possible and important.

Chapter 3: Logical Empiricism and the Rejection of Discovery as Philosophical

3.1 The Logical Empiricist Program

The logical empiricists (逻辑经验主义者) of the Vienna Circle — including Rudolf Carnap, Moritz Schlick, and Otto Neurath — sought to reconstruct scientific knowledge on the basis of logical analysis and empirical observation alone. Their program had two central pillars: first, all meaningful statements are either analytically true (by virtue of meaning) or empirically verifiable (in principle); second, the logical structure of scientific theories can be made explicit through formalization.

Within this framework, the Reichenbachian verdict against the philosophical study of discovery was not merely an incidental aside but a consequence of the broader program. If philosophy’s job is logical analysis, and discovery involves psychological processes not susceptible to logical analysis, then discovery falls outside philosophy’s domain.

3.2 Carnap and the Logic of Confirmation

Rudolf Carnap’s project of inductive logic (归纳逻辑) represents the most technically sophisticated attempt to formalize the context of justification. Carnap’s aim was to define, for any hypothesis \( H \) and evidence statement \( E \), a real-valued confirmation function \( c(H, E) \) that measured the degree to which \( E \) confirmed \( H \). This function was to be defined purely logically — purely in terms of the formal relationships between the sentences \( H \) and \( E \), without reference to the history of how \( H \) was formed.

Carnap’s program illustrates the logical empiricist ideal: if confirmation can be fully formalized, then the epistemological task of science is complete once the formal relationships are specified. The discovery of \( H \) — however it occurred — is irrelevant to its confirmational standing.

The Carnapian program encountered serious technical difficulties. Carnap’s confirmation functions were defined relative to an interpreted formal language, and different choices of language yielded different confirmation values. More fundamentally, the program could not straightforwardly handle universal generalizations (the confirmation of “all ravens are black” by individual instances remained problematic) or the confirmation of theoretical claims that went beyond direct observational content. These difficulties were not fatal to the program but they suggested that formalization of justification was far harder than the positivist ideal implied.

3.3 Karl Popper and the Method of Conjecture

Karl Popper, though not himself a logical empiricist, reinforced the exclusion of discovery from philosophy by a different route. In The Logic of Scientific Discovery (1934/1959), Popper argued that there is no logical method for generating scientific hypotheses. Hypotheses arise through creative conjecture — an irreducibly psychological act — and are then submitted to rigorous attempts at falsification. Scientific rationality is entirely a matter of the critical, deductive testing of hypotheses; the genesis of the hypotheses is epistemologically irrelevant.

Popper was explicit: “The question how it happens that a new idea occurs to a man — whether it is a tune, a dramatic conflict, or a scientific theory — may be of great interest to empirical psychology; but it is irrelevant to the logical analysis of scientific knowledge.”

Popper’s position is notable because the title of his major work — The Logic of Scientific Discovery — might suggest a positive theory of how hypotheses are generated. But Popper’s “logic” is entirely on the justification side: it is a logic of testing and refutation, not of generation. The title is deliberately ironic: there is no logic of discovery in the sense of hypothesis generation, only a logic of falsification.

3.4 Popper’s Deeper Argument: Against Inductivism

Popper’s dismissal of discovery was connected to his deeper attack on inductivism. The standard inductivist picture held that scientists first observe, then generalize: from many observations of white swans, the scientist infers that all swans are white. On this picture, the method of discovery is induction — the generalization from observed cases to universal laws.

Popper argued that this inductivist picture was doubly wrong. First, induction is logically invalid: no finite number of positive instances can establish a universal claim. Second, and more relevant to our purposes, scientists do not actually proceed by first observing and then generalizing. They begin with a problem, propose a bold conjecture, and then subject it to testing. Observation is always selective and theory-directed, not neutral and exhaustive.

This anti-inductivist argument has genuine force, but it leaves a gap: if scientists begin with conjectures, what guides them in forming particular conjectures rather than others? Popper’s answer — that this is a psychological matter beyond rational analysis — is precisely what later philosophers of discovery contested.

3.5 The Legacy and Limits of the Positivist Consensus

The positivist consensus on discovery had a lasting institutional effect: for several decades, philosophy of science journals rarely published papers on discovery, and graduate programs rarely taught the topic as a substantive area of inquiry. The philosophy of science was identified almost entirely with confirmation theory, the structure of explanation, and the theory-observation distinction.

The rehabilitation of discovery that began in the 1960s and accelerated in the 1970s was driven partly by the growing influence of historians of science — above all Kuhn — who showed that the actual development of science could not be understood without attending to the processes by which new ideas and frameworks were generated. If philosophy of science was to remain connected to scientific practice, it could not continue to ignore discovery.

A second driver was the rise of cognitive science and artificial intelligence. When Simon and his colleagues showed that computational programs could rediscover empirical laws by applying explicit heuristics, the positivist claim that discovery processes have no rational structure became empirically untenable. Something with rational structure — computable heuristic search — was producing something recognizably similar to scientific discovery. This did not prove that all discovery reduces to heuristic search, but it demonstrated that the blanket exclusion of discovery from rational analysis was unwarranted.

A third driver, less often acknowledged, was the political and sociological turn in science studies. Sociologists of scientific knowledge (Barry Barnes, David Bloor, Harry Collins) argued that even the context of justification was socially conditioned, blurring the boundary from the other direction. If justification itself is socially shaped, then the privileged status of the context of justification — the domain that the logical empiricists had reserved for rational analysis — was no longer secure. This provoked a renewed interest in whether rational analysis could be extended to discovery to compensate for its loss of exclusive jurisdiction over justification.

Chapter 4: Hanson — Observation, Theory-Ladenness, and Abduction

4.1 The Theory-Ladenness of Observation

N.R. Hanson’s Patterns of Discovery (1958) is one of the foundational texts in post-positivist philosophy of science. Hanson’s central claim is that observation is theory-laden (观察是理论负载的): what a scientist sees when examining a phenomenon is not a raw, neutral sensory given, but is already structured by the theoretical concepts and expectations that the observer brings to the encounter.

Hanson illustrates this with a famous thought experiment. Suppose Tycho Brahe and Johannes Kepler are both watching the dawn. Do they see the same thing? In one sense, yes: both receive the same retinal stimulation as the sun appears over the horizon. But in another, deeper sense, they see something quite different. Tycho sees the sun moving relative to a stationary earth; Kepler sees the earth rotating such that the horizon drops to reveal the stationary sun. The same visual experience is organized under incompatible theoretical frameworks. For Hanson, the philosophically significant sense of “see” — what he calls seeing that (看见) — is already theory-laden.

4.2 Implications for the Discovery/Justification Distinction

If observation is theory-laden, the clean separation between context of discovery and context of justification becomes harder to maintain. The logical empiricist picture assumed that the evidence for a hypothesis consisted of observation reports that could be specified in a theory-neutral observation language. But if no such language is available — if all observation reports are already saturated with theoretical content — then the “evidence” available in the context of justification is not independent of the theoretical commitments active in the context of discovery.

This has important consequences. It means that scientists working within different theoretical frameworks may not simply disagree about which theory is best supported by neutral evidence; they may also disagree about what the evidence is. The problem of incommensurability (不可通约性) — most fully developed by Kuhn — has its philosophical roots in Hanson’s observation thesis.

4.3 Peirce’s Three-Way Distinction

Before developing Hanson’s account of abduction, it is important to appreciate how carefully Peirce distinguished abduction from the other two inferential forms, because subsequent discussions have frequently blurred this distinction in ways that distort the philosophical issues.

Peirce worked out his three-way distinction over several decades, and his formulations shifted. In his mature view, the three forms differ not merely in their logical structure but in their epistemic function and their direction of inference.

Deduction proceeds from a rule and a case to a result. It is ampliative in the sense that it is necessary — the conclusion must be true if the premises are true — but it is not ampliative in the epistemic sense: the conclusion contains only what the premises already contain, explicitly or implicitly. Deduction is the inferential form of mathematics and formal logic. An example: All copper conducts electricity (rule); this wire is copper (case); therefore, this wire conducts electricity (result).

Induction proceeds from cases and results to a rule. It is epistemically ampliative — the conclusion goes beyond what is contained in the premises — and therefore uncertain. It is the inferential form appropriate to statistical generalization and hypothesis confirmation. An example: these fifty copper wires all conduct electricity (cases and results); therefore, all copper conducts electricity (rule). Induction is directed at establishing the generality of a regularity already observed in particular cases.

Abduction (which Peirce also called retroduction or, in his later writings, hypothesis) proceeds from a rule and a result to a case. It is the inferential form appropriate to the generation of explanatory hypotheses. An example: Electrical conductors expand when current passes through them (rule); this wire has expanded (result); therefore, current may have passed through this wire (case — i.e., hypothesis). Crucially, the conclusion of an abductive inference is not established by the inference; it is merely nominated as worthy of investigation. Peirce was insistent on this point: abduction does not confirm a hypothesis but selects it for further testing.

The distinctiveness of Peirce’s abduction lies in what he called the “surprising fact” condition. Abduction is triggered not by any gap in knowledge but by phenomena that are anomalous — that resist assimilation to the current theoretical framework. The surprising character of the result is what licenses the generation of a new hypothesis.

Peirce also insisted that abduction has a formal structure that distinguishes it from mere guessing: the abduced hypothesis must be such that, if true, it would explain the surprising result — would render it a matter of course rather than a surprise. This explanatory condition constrains abduction and distinguishes good abductive hypotheses from bad ones.

4.4 Abduction as the Logic of Discovery

Hanson’s second major contribution is his defense of abduction (溯因推理) as a genuine logical pattern operative in the context of discovery. Drawing on Peirce, Hanson argued that the logical empiricists and Popper had given a distorted picture of scientific inference because they recognized only two inferential forms — deduction and induction — and therefore had no logical form to assign to the generation of hypotheses.

Abduction is not a deductively valid inference form — the conclusion does not follow necessarily from the premises. But Peirce argued it is not irrational either. It is the inference form by which scientists generate hypotheses worth investigating. The abductive form licenses the proposal of a hypothesis not because the hypothesis is proven, but because it would render intelligible what is otherwise anomalous.

Hanson developed this account of abduction with careful attention to the history of science, arguing that the reasoning of Kepler as he arrived at his laws of planetary motion instantiated this pattern. Kepler did not simply guess; nor did he mechanically generalize from data. He reasoned: the observed orbital data is anomalous given circular orbits; but if orbits are elliptical with the sun at one focus, the data follows as a matter of course; therefore, elliptical orbits are a serious hypothesis.

4.5 Inference to the Best Explanation: Lipton’s Development

Peter Lipton’s Inference to the Best Explanation (2004) represents the most systematic philosophical development of the abductive tradition. Lipton argues that the form of inference central to science is not deduction, not Humean induction, but what he calls inference to the best explanation (最佳解释推断, IBE).

The core thesis of IBE is that scientists — and ordinary reasoners — infer the truth of the hypothesis that would, if true, provide the best explanation of the available evidence. The inference is licensed not by the mere fact that the hypothesis explains the evidence, but by the comparative judgment that it explains the evidence better than any available competitor.

Lipton introduces a crucial distinction between the likeliest explanation and the loveliest explanation. The likeliest explanation is the one most probably true given the evidence. The loveliest explanation is the one that would, if true, provide the deepest and most illuminating understanding. Lipton’s controversial thesis is that scientists track loveliness rather than likeliness: they prefer the hypothesis that would be most explanatorily illuminating, and this preference is not merely an aesthetic preference but an epistemically reliable guide to truth.

This thesis is contentious. Critics object that there is no guarantee that the universe is organized to reward our preferences for lovely explanations. The pragmatic success of science might reflect that loveliness and likeliness are generally correlated — perhaps because highly explanatory theories tend to be simpler, more general, and less ad hoc — but this correlation is not logically necessary. Bas van Fraassen has pressed this point as part of his broader empiricist critique of inference to the best explanation.

4.6 Critiques of Abduction

Abduction and IBE face several serious objections that must be reckoned with.

The underconsideration problem: IBE tells us to infer the best explanation among those considered. But if the true explanation is not among the hypotheses we have generated, IBE will lead us to accept a false hypothesis as the best available. The inference is only as good as the pool of candidates. This objection shows that abduction is not a self-sufficient logic of discovery; it must be supplemented by an account of how we generate the relevant pool of candidate hypotheses.

The no-miracles and base-rate problem: Van Fraassen argues that IBE licenses a kind of circular reasoning: we appeal to the explanatory success of science to justify the reliability of IBE, but this presupposes that the best explanations are true — which is exactly what IBE claims. The argument from scientific success to the truth of IBE is itself an inference to the best explanation, making the defense circular.

The competing hypotheses problem: In many scientific contexts, multiple incompatible hypotheses explain the available evidence equally well. IBE has no principled mechanism for selecting among equally explanatory competitors. Some philosophers (Thagard) have proposed that criteria such as simplicity, explanatory breadth, and non-ad-hocness can break ties, but it remains contested whether these meta-criteria have genuine epistemic force.

4.7 Evaluating Hanson’s Account

Hanson’s account raises important questions. If abduction is a genuine logical form, what distinguishes good abductive inferences from bad ones? Any number of hypotheses might, if true, render a given anomaly intelligible; what makes some abductive hypotheses better than others? Hanson’s account is most powerful as a critique of the positivist exclusion of discovery from logic; it is less fully developed as a positive theory of what makes a discovery good.

The tradition from Peirce through Hanson to Lipton has substantially enriched our understanding of the inferential structure of hypothesis generation. The key achievement is to show that the generation of hypotheses is not arbitrary: it is constrained by the explanatory requirement, by comparative assessment of candidate explanations, and by background theoretical commitments. Whether this rich structure constitutes a genuine logic of discovery — in the sense of an inference form with sufficient normative strength to guide practice — remains contested.

Chapter 5: Kuhn on Discovery — Anomaly, Crisis, and Gestalt Shift

5.1 The Structure of Scientific Revolutions

Thomas Kuhn’s The Structure of Scientific Revolutions (1962) transformed the philosophy of science by replacing the logical empiricists’ static image of science with a dynamic, historical account of how science actually develops. For our purposes, Kuhn’s account is important because it provides what Reichenbach explicitly denied: a philosophically structured account of how major theoretical discoveries come about.

Kuhn identifies a recurring pattern in the history of science. Normal science operates within a paradigm (范式) — a set of exemplary solved problems, shared theoretical commitments, methodological rules, and standards of what counts as a good scientific problem and solution. Normal science is not aimed at discovery in any dramatic sense; it is “mopping up” — working out the implications of the paradigm, applying it to new domains, and improving its precision.

5.2 Paradigms and Normal Science

The concept of a paradigm is central to Kuhn’s philosophy of science, and it is more complex than casual usage suggests. Kuhn distinguished at least two senses of “paradigm” in his later writing: the disciplinary matrix (the whole constellation of beliefs, values, and techniques shared by a scientific community) and the exemplar (the concrete problem solutions that the community uses as models for tackling new problems).

The exemplar sense is particularly important for understanding discovery. Scientists learn to do normal science not by memorizing explicit rules but by working through canonical problem solutions and developing the capacity to recognize new problems as structurally similar to previously solved ones. This is a form of tacit knowledge — the ability to see a new problem as an instance of a familiar type — and it is not easily articulable in formal rules.

During normal science, anomalies are not absent; they are ubiquitous. But normal science has well-developed mechanisms for managing anomalies: setting them aside for future resolution, reclassifying them as measurement errors, or making local adjustments to the paradigm that preserve its core commitments. The crucial question is what transforms an ordinary anomaly into a crisis-inducing one.

Kuhn’s answer is that some anomalies strike at the fundamental commitments of the paradigm — at the exemplary problem solutions that have defined what normal science is. When such anomalies persist despite repeated attempts at resolution, when they attract the sustained attention of leading scientists, and when attempts at resolution themselves multiply into a proliferation of competing adjustments that collectively undermine the paradigm’s authority, crisis ensues.

5.3 Anomaly and Crisis

Discovery of a fundamentally new kind becomes possible only when anomaly (异常) is recognized. An anomaly is a phenomenon that resists assimilation to the existing paradigm despite persistent effort. Not every failed prediction constitutes an anomaly in this sense; scientists routinely set aside recalcitrant data as measurement error or as a problem to be solved later. An anomaly becomes philosophically significant when it persists despite sustained attempts to accommodate it within the paradigm.

Kuhn emphasizes that the transition from anomaly-recognition to crisis is not automatic. It requires the anomaly to be taken seriously, to attract the sustained attention of leading scientists, and to resist repeated attempts at resolution. When crisis arrives, the normal rules of science loosen: scientists become willing to entertain speculative hypotheses that would have been rejected as insufficiently paradigm-conforming during normal science.

5.4 Discovery as Gestalt Shift

Kuhn’s most provocative claim is that revolutionary discovery involves something like a gestalt shift (格式塔转换) — a sudden reorganization of the entire conceptual field that cannot be fully explained as a logical inference from prior evidence. Drawing on the psychology of visual perception, Kuhn compares the scientist’s transition from one paradigm to another to the experience of suddenly seeing a Necker cube flip or seeing the duck-rabbit figure switch from duck to rabbit. There is a before and after, but no rational reconstruction of the steps in between.

5.5 The Rationality Objection

Kuhn’s account of revolutionary discovery attracted fierce criticism on grounds of rationality, and addressing this objection is essential for evaluating the philosophical legacy of The Structure of Scientific Revolutions.

The rationality objection has two versions. The first, weaker version holds that Kuhn’s account of gestalt shifts and incommensurability makes paradigm change appear irrational — a conversion experience rather than a reasoned response to evidence. If scientists cannot fully translate one paradigm’s claims into another’s terms, and if the standards of evidence are paradigm-relative, then it seems that paradigm change cannot be rationally justified and must instead be explained sociologically or psychologically.

Kuhn’s response to this weaker objection is that he was misread. He never claimed that paradigm change is irrational; he claimed that it is not logically compelled by a neutral body of evidence. The difference is important. A choice can be rational — supported by good reasons — without being logically forced. Kuhn identified a set of values that scientists share across paradigms and that govern theory choice: accuracy (the theory should agree with experiment), consistency (internal and with established theories), breadth (the theory should explain a wide range of phenomena), simplicity (the theory should be parsimonious), and fruitfulness (the theory should suggest new discoveries). These values are shared even when they do not uniquely determine a verdict, and their shared character is what makes paradigm change a rationally assessable process rather than a mere conversion.

The second, stronger version of the rationality objection presses the issue further. Even if Kuhn’s five values are shared across paradigms, they underdetermine theory choice: two equally rational scientists can weigh the values differently and reach different verdicts. This means that paradigm choice is not rationally determined, which raises the question of what explains the actual pattern of paradigm adoption in historical cases. If the rational factors underdetermine the choice, are the remaining determinants non-rational?

Kuhn’s response was to acknowledge that the rational factors underdetermine theory choice and to insist that this is not a defect of science but a feature. Science is not an algorithm; it requires the exercise of judgment, and rational individuals exercising judgment under genuine uncertainty may disagree. What matters for scientific progress is that the community as a whole, applying shared values with individual diversity of weighting, reliably converges on empirically successful theories over time. The rationality of science is a community-level property, not simply a property of individual choices.

5.6 Philosophical Implications of Kuhn’s Account

Kuhn’s account of discovery has several important philosophical implications.

First, it suggests that discovery is not a moment but a process — extended over time, involving multiple stages (anomaly recognition, crisis, paradigm shift, community acceptance), and not reducible to a single inferential step.

Second, it raises the problem of incommensurability: if different paradigms organize observation and evidence differently, it may be impossible to fully translate the claims of one paradigm into the terms of another. This challenges the idea that there is a neutral evidential base on which competing theories can be rationally compared.

Third, Kuhn’s account raises questions about the rationality of revolutionary discovery that generated an extensive philosophical debate. As discussed, Kuhn ultimately defended a value-based account of theory choice that is rational without being algorithmically deterministic.

Fourth, Kuhn’s account has important implications for scientific education. If paradigms are transmitted through exemplars rather than explicit rules, and if normal science consists in extending exemplars to new cases, then the education of scientists is not primarily a matter of teaching them methodological rules but of training them to see new problems through the lens of canonical solutions. The tacit dimension of paradigm-based training connects Kuhn’s account to Polanyi’s epistemology of tacit knowledge.

Chapter 6: Historical Case Studies — Benzene, X-rays, DNA

6.1 The Role of Case Studies in Philosophy of Discovery

Historical case studies serve a critical function in the philosophy of scientific discovery. They supply the empirical base against which philosophical theories of discovery can be tested. A theory of discovery that has no contact with the actual history of science risks becoming merely a priori speculation. Conversely, a historical account that has no philosophical framework risks becoming mere narrative.

The case studies examined in this chapter — Kekulé’s discovery of the benzene ring, Röntgen’s discovery of X-rays, Fleming’s discovery of penicillin, and Watson and Crick’s determination of the DNA structure — are philosophically rich because they instantiate different discovery patterns and raise different epistemological issues. They collectively provide what might be called a natural experiment in discovery methodology: diverse modes of inquiry, diverse fields, diverse outcomes, but all recognizably scientific.

One methodological caution is in order. The historical record of scientific discoveries is frequently distorted by retrospective reconstruction. Scientists, writing after their discoveries, often present a more orderly and logically compelling narrative than the actual process warranted. Kekulé’s famous account of his dream is probably the most discussed example of this phenomenon. Historians of science have learned to treat scientists’ retrospective accounts with critical skepticism and to supplement them with contemporary documents — laboratory notebooks, correspondence, and published preliminary reports — that preserve the actual texture of the discovery process.

6.2 Kekulé and the Benzene Ring (1865)

August Kekulé’s proposal of the hexagonal ring structure for benzene in 1865 is one of the most famous examples of creative discovery in the history of chemistry. Kekulé himself famously reported arriving at the ring structure through a dream or hypnagogic vision in which he saw a snake seizing its own tail.

Philosophically, Kekulé’s case is significant for several reasons. It seems to exemplify the role of analogical reasoning (类比推理) in discovery — the snake image (whether literal or metaphorical) functions as a structural analogy that suggested a new form of molecular organization. It also raises questions about the relationship between creative intuition and rational assessment: Kekulé proposed the ring structure on the basis of a vision, but the structure was subsequently supported by extensive experimental evidence. Does the mode of generation affect the epistemic status of the discovery?

On the Reichenbachian view, it does not: the ring structure’s epistemic status is determined by the evidence for it, not by the dream that suggested it. But the case invites reflection on a subtler point: perhaps creative methods of hypothesis generation — including dreams, analogies, and visual imagery — are epistemically valuable precisely because they escape the constraints of the prevailing theoretical framework and allow conceptual reorganization that more rule-governed methods would block. If so, the irrationality of the generation process is not merely tolerated but is productively necessary.

6.3 Röntgen and X-rays (1895)

Wilhelm Röntgen’s discovery of X-rays in November 1895 is a paradigmatic case of serendipitous discovery (偶然发现). Röntgen was investigating cathode ray tubes when he noticed that a barium platinocyanide screen across the room was fluorescing, even though it was shielded from any direct radiation the tube might emit. He inferred that the tube was emitting a previously unknown form of radiation capable of penetrating opaque materials.

Röntgen’s case illustrates the importance of prepared attention — what Louis Pasteur called the principle that chance favors the prepared mind. Röntgen noticed the anomalous fluorescence precisely because he was an expert investigator who recognized that it required explanation. A less informed observer might have dismissed it as irrelevant.

6.4 Fleming and Penicillin (1928)

Alexander Fleming’s discovery of penicillin in 1928 shares with Röntgen’s discovery the element of accident: Fleming noticed that a mold contaminant (Penicillium notatum) had cleared a zone of bacterial growth on a petri dish. But the philosophical interest of the case lies in what Fleming did with this observation.

Fleming recognized the zone of clearing as anomalous — not a failure of experimental technique but an indication that the mold was producing a substance with antibacterial properties. This recognition instantiated an abductive inference: the most natural explanation of the clearing was that the mold was secreting a bactericidal agent. Fleming pursued this hypothesis, named the putative agent penicillin, and documented its antibacterial spectrum — though full therapeutic development required the subsequent work of Florey and Chain.

The Fleming case illustrates the multi-stage character of discovery. The initial observation, the abductive inference, the preliminary experimental confirmation, and the eventual therapeutic development were spread across decades and involved multiple scientists and disciplines. Any account that locates “the discovery” at a single moment — Fleming’s observation of the contaminated plate — is philosophically misleading. Discovery, in the philosophically interesting sense, is a process.

6.5 Watson and Crick and the Structure of DNA (1953)

The determination of the double-helix structure of DNA by James Watson and Francis Crick in 1953 is philosophically the richest of our case studies, because it exemplifies a distinctive discovery methodology — constraint-based model building — that is neither straightforwardly abductive nor inductively driven.

Watson and Crick’s approach was guided by constraints from multiple independent sources. Rosalind Franklin’s X-ray crystallography data — most critically, the famous Photo 51 obtained by Franklin and her student Raymond Gosling — established the helical geometry of DNA and constrained the backbone dimensions. Erwin Chargaff’s biochemical rules established that in any DNA sample, the proportion of adenine equals the proportion of thymine, and the proportion of guanine equals the proportion of cytosine (A=T, G=C). Jerry Donohue’s expertise in physical chemistry corrected Watson’s initial error in the tautomeric forms of the bases, enabling the correct base-pair geometry. The biological requirement that the molecule must be capable of faithful replication added a further functional constraint.

This case illustrates what philosophers of science call consilience of inductions — the convergence of independent lines of evidence on a single hypothesis as a particularly powerful form of evidential support. The convergence is philosophically important because it provides stronger evidential support than any single line of evidence could. When independent methods, employing different experimental techniques and background theories, all point to the same structural model, the probability that the model is merely an artifact of one technique’s systematic error is greatly reduced.

The DNA case also raises questions about the ethics of scientific discovery that connect epistemology to the sociology and ethics of science. Franklin’s X-ray data were used without her knowledge or full consent — Watson and Crick had access to Photo 51 through Wilkins, without Franklin’s awareness that her data were being shared. This raises a question about the relationship between epistemic and moral dimensions of discovery: does the wrongful acquisition of data affect the epistemic status of the conclusions derived from it? The epistemic and ethical dimensions are logically independent — the double helix is the correct structure regardless of how the data were obtained — but the case illustrates that the social processes of discovery have moral dimensions that cannot be ignored in a complete philosophical account.

6.6 What the Case Studies Teach Us Philosophically

Taken together, the four case studies support several philosophical conclusions.

First, discovery is genuinely diverse in its epistemic structure. Kekulé’s case highlights the role of creative analogy and visual imagination; Röntgen’s case highlights serendipity and prepared attention; Fleming’s case highlights abductive inference from anomaly; Watson and Crick’s case highlights constraint-based model building and consilience. No single inferential pattern captures all cases. This supports the pluralist conclusion advanced in Chapter 9.

Second, discovery is typically multi-stage and collaborative. Even in cases that are retrospectively narrated as individual insights — Kekulé’s dream, Fleming’s contaminated plate — the discovery process involves extended investigation, multiple contributors, and gradual community acceptance. The philosophical focus on the “moment of discovery” is often misleading.

Third, context and expertise matter enormously. Röntgen and Fleming both noticed things that others had either not seen or not taken seriously. Their expertise — their theoretical preparation and experimental skill — is what made the anomalies visible as scientifically significant. This supports Polanyi’s point about the tacit dimension of discovery: expert perception of significance is not reducible to the application of explicit rules.

Fourth, the ethical and social dimensions of discovery are philosophically inseparable from its purely epistemic dimensions. The credit allocation in the DNA case, the institutional arrangements that gave Watson access to Franklin’s data, and the differential reception of scientists by their communities are not merely sociological background to the epistemological story; they affect what gets discovered, by whom, and when.

Chapter 7: Computational and Heuristic Models of Discovery (Simon)

7.1 Simon’s Research Program

Herbert Simon’s Models of Discovery (1977) represents the most sustained attempt to develop a rigorous, computational account of scientific discovery. Simon, a Nobel laureate in economics and a pioneer of artificial intelligence, argued that the processes of scientific discovery could be modeled as heuristic search (启发式搜索) through a problem space.

Simon’s central claim is that discovery is a form of problem solving, and that problem solving — even creative problem solving — can be characterized by computational processes that are in principle formalizable. This does not mean that discovery is mechanical or deterministic; heuristic search involves selective, non-exhaustive exploration of possibility spaces guided by evaluation functions that assess progress toward a goal.

Simon’s research program grew out of his earlier work with Allen Newell on the General Problem Solver (GPS), a computer program designed to model human problem solving in terms of means-ends analysis. GPS operated by identifying the difference between the current state and the goal state and applying operators that reduced the difference. Simon extended this framework to scientific discovery, arguing that scientific problem solving is a special case of the general problem-solving schema.

7.2 BACON and Computational Discovery Programs

Simon and his colleagues developed a family of computer programs — most notably the BACON program — that were designed to rediscover empirical laws from data. BACON operated by searching for regularities in numerical data, guided by heuristics such as “if two variables covary, form their ratio and look for invariances.”

BACON successfully rediscovered several important scientific laws, including Kepler’s third law of planetary motion (that the square of a planet’s period is proportional to the cube of its orbital radius), Ohm’s law relating voltage, current, and resistance, and the ideal gas law. These rediscoveries were taken as evidence that the inferential processes of discovery have computational structure.

BACON was followed by a family of successor programs: DALTON (which could rediscover the atomic weight ratios in chemical compounds), STAHL (which modeled qualitative inference about chemical reactions), and GLAUBER (which induced qualitative laws relating chemical substances). This family of programs demonstrated that the BACON approach could be extended beyond purely quantitative law discovery to qualitative and structural domains.

The philosopher of science Paul Thagard developed a complementary computational approach in his program ECHO, which modeled the evaluation of competing scientific hypotheses using a connectionist network that tracked explanatory coherence. ECHO could simulate the pattern of theory acceptance and rejection in several historical episodes, including the Lavoisier oxygen case. This work extended Simon’s computational approach from data-driven discovery to theory-level inference.

7.3 The Significance of Computational Discovery

Simon’s claim that computational programs can model scientific discovery is philosophically significant because it directly challenges the positivist assumption that discovery is cognitively inaccessible to rational analysis. If a program running explicit rules can produce something recognizably similar to scientific discovery, then the discovery process has structure that is, at least in principle, capturable by formal methods.

Simon was careful to note that the BACON programs did not claim to model the full richness of scientific discovery. They modeled one component — the data-driven induction of quantitative laws — within a broader cognitive process that also includes experiment design, concept formation, theoretical integration, and the recognition of significance. Simon viewed BACON as a proof of concept: a demonstration that at least one component of discovery has computable structure.

The programs also have a pedagogical lesson: they make explicit what heuristics are operative in a particular type of discovery, and this explicitness allows those heuristics to be evaluated, compared, and potentially improved. This is a form of rational control over the discovery process that the positivist picture denied was possible.

7.4 Limitations and Objections

Simon’s computational approach to discovery has attracted several important criticisms.

The representation problem: The BACON program begins with data already organized in numerical form, with variables already identified and measured. In actual scientific discovery, a major challenge is determining what to measure and how to represent the phenomena. The choice of representation is not itself a heuristic search over a pre-given problem space — it is itself a creative theoretical act. Lavoisier did not simply apply heuristics to pre-existing data; he redesigned the very apparatus of chemistry, replacing qualitative descriptions of burning with quantitative measurement of mass. The decision to measure mass precisely was a conceptual innovation that preceded the discovery of oxygen, not a consequence of applying standard heuristics.

The theory-generation problem: BACON discovers empirical regularities, but major scientific discoveries often involve the postulation of theoretical entities (理论实体) — electrons, genes, fields — that are not directly observable. Heuristic search over observational data cannot by itself generate theoretical-level discoveries. A program that can rediscover Kepler’s third law cannot, without radical extension, generate the Newtonian gravitational theory that explains why Kepler’s law holds.

The meaning of “discovery”: Critics have questioned whether BACON genuinely discovers laws or merely detects patterns. Discovery seems to require understanding — grasping why a regularity holds, not merely that it holds. Computational pattern detection may fall short of the epistemic achievement constitutive of genuine discovery. Simon’s response was that “understanding” is itself a cognitive state that could, in principle, be given a computational characterization; the objection assumes that understanding is irreducibly non-computational, which is a substantive and controversial claim.

The ecological validity problem: BACON is supplied with clean, preselected data. Real scientific data are noisy, incomplete, and embedded in a web of experimental context that must be interpreted before the data can be used. The gap between laboratory data and the organized numerical tables that BACON receives as input involves substantial scientific work that the BACON model does not represent.

Simon acknowledged these limitations and viewed the BACON program not as a complete theory of discovery but as a model of one component — data-driven law induction — within a broader account that would need to address theoretical inference as well.

7.5 Contemporary Extensions: Machine Learning and Discovery

Simon’s work has found unexpected vindication in contemporary machine learning. Systems capable of discovering empirical patterns from large datasets — from protein structure prediction (AlphaFold) to the identification of novel mathematical conjectures (DeepMind’s work with mathematicians) — demonstrate that computational methods can produce genuinely novel outputs that were not predictable from the programs’ explicit rules. This raises fresh versions of the philosophical questions Simon’s work posed: is this discovery? does it require understanding? and can it illuminate the human discovery process?

These questions have become urgently practical as well as philosophical. If computational systems can perform aspects of scientific discovery, this has implications for the organization of research, the allocation of credit, and the very conception of scientific expertise.

Chapter 8: Analogical Reasoning and Discovery

8.1 The Structure of Analogical Inference

Analogical reasoning (类比推理) is one of the most pervasive cognitive strategies in scientific discovery. It proceeds by identifying structural similarities between a better-understood source domain (源域) and a less-understood target domain (目标域), and using the structure of the source domain to generate hypotheses about the target.

The inferential strength of an analogical argument depends on several factors: the number and variety of shared properties, the relevance of the shared properties to the inferred property, the degree to which the two domains are structurally similar at a deep level, and the absence of relevant disanalogies.

Mary Hesse, in her work on models and analogies in science, distinguished positive analogies (properties the source and target are known to share), negative analogies (properties known to differ), and neutral analogies (properties of the source whose status in the target is unknown). The epistemically productive region of an analogy lies in the neutral analogies: these are the properties that the analogy suggests investigating in the target domain, and the discovery process consists in turning neutral analogies into positive or negative ones through experiment and observation.

8.2 Maxwell’s Electromagnetic Theory

James Clerk Maxwell’s development of electromagnetic field theory in the 1860s is one of the most philosophically instructive episodes of analogical discovery in the history of physics. Maxwell did not begin with electromagnetic phenomena alone; he began with an explicit mechanical analogy.

In his early paper “On Faraday’s Lines of Force” (1855–56), Maxwell modeled the electromagnetic field on the flow of an incompressible fluid through a porous medium. Faraday had described electromagnetic phenomena in terms of “lines of force” — a qualitative, geometric description. Maxwell’s strategy was to find a physical system whose mathematical description matched Faraday’s qualitative picture, then use the mathematics of that system as a scaffold for analyzing electromagnetic phenomena.

The fluid analogy worked as follows. Faraday’s electric field lines were modeled as the streamlines of an incompressible fluid. The density of the fluid corresponded to the intensity of the electric field; the velocity of the fluid corresponded to the direction of the field. Maxwell could then use the well-developed mathematics of fluid dynamics — the equations of Laplace and Stokes — to derive mathematical relations among electromagnetic quantities.

The crucial philosophical point is that Maxwell did not believe the electromagnetic field was a fluid. He used the fluid analogy as a mathematical scaffold — a way of borrowing mathematical structure from a well-understood domain to generate equations in a less-understood one. Once the equations were established, their physical interpretation could be developed independently of the original analogy.

In his later paper “On Physical Lines of Force” (1861–62), Maxwell introduced a different analogy: a mechanical model of rotating vortices and idle wheels that represented the electromagnetic medium. This more elaborate mechanical model generated new predictions — including a displacement current — that Maxwell had not anticipated from electromagnetic data alone. The prediction that electromagnetic disturbances propagate as waves, at a speed calculable from electric and magnetic constants, and that this speed matched the measured speed of light, was the fruit of analogical reasoning from a mechanical model.

The Maxwell case illustrates several philosophical points about analogical discovery. First, analogies can function as generative mechanisms: the mechanical model not only organized existing data but generated genuinely new predictions that were not contained in the electromagnetic data themselves. Second, analogies need not be literally true to be epistemically productive. Maxwell abandoned the mechanical models in the Treatise on Electricity and Magnetism (1873), presenting the electromagnetic field equations on their own terms. The analogies had served their heuristic purpose; they were scaffolding to be removed once the building could stand on its own. Third, the history of Maxwell’s work illustrates how multiple analogies, applied in sequence, can each contribute to a discovery process: the fluid analogy organized Faraday’s qualitative description; the vortex analogy generated new predictions; and the final abstract formulation transcended both.

8.3 Huygens and the Wave Theory of Light

Christian Huygens’s development of the wave theory of light in the seventeenth century provides an earlier illustration of analogical discovery and one that raises distinct philosophical issues. Huygens proposed that light, like sound, propagates as a wave through a medium. This analogy with sound waves guided his investigation of optical phenomena.

The analogy suggested several predictions that were subsequently investigated. If light is a wave, it should exhibit diffraction — bending around obstacles — just as sound waves do. It should also exhibit interference — the characteristic pattern of constructive and destructive superposition — that is the signature of wave behavior. Huygens’s principle, according to which every point on a wavefront can be treated as a source of secondary wavelets, was directly analogical with the mechanism of sound propagation.

The wave theory faced a significant difficulty: waves of sound require a medium (air) through which to propagate. If light is a wave, it requires a medium too — the luminiferous ether (以太). The ether hypothesis generated a major research program spanning two centuries, culminating in the Michelson-Morley experiment (1887) that failed to detect any evidence of the ether. This failure ultimately contributed to the development of special relativity by Einstein in 1905.

The Huygens case illustrates a philosophically important feature of analogical reasoning in discovery: analogies can generate false as well as true predictions. The ether hypothesis, motivated by the analogy between light and sound, was a productive research program that ultimately led to its own refutation. The productivity of the analogy — the research it generated, the precision it demanded, the anomalies it revealed — was epistemically valuable even though one of its central commitments (the ether) was false. This suggests that the epistemic value of analogical reasoning in discovery cannot be measured simply by whether the analogy’s predictions are true; it must also be measured by the quality of the research program the analogy generates.

8.4 Darwin and Artificial Selection

Charles Darwin’s development of the theory of natural selection provides one more instructive example of analogical discovery, this time in biology. Darwin was struck by the efficiency with which animal breeders could modify the traits of domesticated species over relatively few generations by systematically selecting for desirable characteristics. The analogy between artificial selection (人工选择) and natural processes formed a central scaffold for his theorizing.

The analogy worked as follows. Breeders select parent organisms with desirable traits; their offspring tend to inherit those traits; over many generations, the trait becomes fixed in the population. Nature “selects” not by breeders’ preferences but by differential reproductive success: organisms whose heritable traits make them better suited to their environment reproduce more prolifically, and their traits spread through the population.

What the analogy provided was not merely a metaphor but a causal mechanism. Darwin could point to artificial selection as a known, observable process that produced exactly the kind of divergence and specialization that the fossil record suggested had occurred in nature. The analogy gave the natural selection hypothesis credibility precisely because it connected a novel and unobserved mechanism to a familiar and well-documented one.

The analogy also guided Darwin’s investigation in a normative sense: it told him what to look for. If natural selection is analogous to artificial selection, he should expect to find (a) heritable variation in populations, (b) differential reproductive success correlated with traits, and (c) gradual directional change in populations over time. All three were confirmed, though the mechanisms of inheritance were not clarified until the development of Mendelian genetics decades later.

8.5 Tacit Knowledge and the Limits of Explicit Method

Michael Polanyi’s account of tacit knowledge (默会知识) provides an important complement to — and qualification of — explicit accounts of analogical and heuristic discovery. Polanyi argued that scientific expertise involves a form of knowing that cannot be fully articulated in explicit rules or procedures. The skilled scientist knows more than she can say: she recognizes promising analogies, fruitful research directions, and significant anomalies through a kind of trained intuition that is not reducible to the application of explicit heuristics.

Polanyi’s famous slogan is “we can know more than we can tell.” This tacit dimension of knowledge has important implications for philosophy of discovery. If significant components of the expert scientist’s capacity for discovery are tacit, then any explicit account of discovery — whether logical, computational, or heuristic — will be incomplete. The limits of explicit methodology may not reflect a deficiency in our philosophical account; they may reflect a genuine feature of the cognitive structure of expertise.

Chapter 9: Can There Be a Logic (or Methodology) of Discovery?

9.1 Restating the Central Question

Having surveyed the major approaches to scientific discovery — the logical empiricist exclusion, Hanson’s abduction, Kuhn’s paradigm shifts, Simon’s heuristic models, analogical reasoning, and Polanyi’s tacit knowledge — we are in a position to return to the central question of the course: Is there a logic or methodology of discovery?

The question can be understood in several ways:

1. The logical question: Is there a formal inferential pattern — like abduction — that characterizes the generation of scientific hypotheses?
2. The methodological question: Are there normative guidelines — heuristics, rules of thumb, or strategies — that reliably improve the prospects of discovery?
3. The evaluative question: Are there criteria by which discovery processes can be assessed as more or less rational, more or less fruitful?

These questions are distinct. A negative answer to the logical question does not entail a negative answer to the methodological or evaluative questions. One might hold that there is no formal logic of discovery — no valid inference form that carries you from evidence to hypothesis — while maintaining that some discovery strategies are demonstrably better than others.

9.2 Arguments Against a Logic of Discovery

Several considerations support the view that there can be no logic of discovery in the strong sense.

The underdetermination of hypotheses by evidence: For any finite body of evidence, infinitely many hypotheses are logically compatible with it. No inferential rule can identify a unique hypothesis as the “correct” product of the evidence. Discovery involves selecting from an infinite space, and logic alone cannot constrain selection to a single candidate.

The creativity objection: The greatest scientific discoveries — general relativity, quantum mechanics, natural selection — seem to involve genuine conceptual creativity that cannot be reconstructed as the application of a pre-given method. Einstein did not derive special relativity by running an algorithm; he engaged in a kind of conceptual experimentation that violated prevailing methodological norms.

The context-sensitivity of discovery: What counts as a fruitful strategy for discovery depends on the current state of the field, the available instrumentation, the theoretical background, and a host of other contextual factors. There may be no domain-general methodology of discovery — only domain-specific heuristics whose fruitfulness is local and provisional.

9.3 Arguments for a Methodology of Discovery

Defenders of a methodology of discovery do not need to claim that there is a logic of discovery in the strong, deductive sense. They need only show that some discovery strategies are systematically more fruitful than others, and that this systematic advantage can be identified, articulated, and taught.

The heuristic evidence: Simon’s BACON programs demonstrate that computational heuristics can successfully rediscover empirical laws. Even if heuristics do not guarantee success, they perform better than random search. This is a methodological asymmetry that warrants investigation.

The abductive evidence: Hanson and Peirce’s account of abduction identifies a genuine inferential pattern — inference to the best explanation — that scientists explicitly employ. The pattern is not deductively valid, but it is not arbitrary either: some abductive hypotheses are better than others by criteria (simplicity, scope, coherence with background theory) that can be articulated.

The historical evidence: The study of past discoveries does reveal recurring strategies — controlled experimentation, systematic variation of parameters, analogical transfer, convergent constraint satisfaction — whose application in new contexts is not irrational. The history of science is not a random walk; it exhibits patterns that can inform future inquiry.

9.4 Methodological Pluralism vs. a Unified Theory of Discovery

The debate between advocates of a unified theory of discovery and advocates of methodological pluralism is one of the most important unresolved issues in the philosophy of scientific discovery.

Advocates of a unified theory hold that beneath the apparent diversity of discovery strategies — abduction, analogical reasoning, heuristic search, constraint satisfaction, paradigm shift — there is a single underlying cognitive or inferential mechanism. Simon came closest to articulating such a theory: all these strategies, he argued, are variants of heuristic search through a problem space. Abduction is a heuristic that selects the explanatorily best candidate from the problem space. Analogical reasoning is a heuristic that maps a known problem space onto an unknown one. Paradigm shifts are large-scale reorganizations of the problem space. If this unified account is correct, then there is a general theory of discovery after all — it is just a theory of heuristic search rather than a theory of logical inference.

The objection to this unified account is that the problem space framework is too abstract to be informative. To say that all discovery is heuristic search through a problem space is like saying that all cognition is information processing — it is trivially true but scientifically and philosophically empty. The interesting questions are about the specific structure of particular problem spaces, the particular heuristics operative in specific discovery contexts, and the specific constraints that govern hypothesis generation in each domain. These domain-specific structures resist unification into a single theory.

Advocates of methodological pluralism hold that the diversity of discovery strategies reflects a genuine diversity in the kinds of epistemic tasks that discovery can involve. Data-driven law discovery calls for inductive heuristics; theoretical hypothesis generation calls for abductive and analogical reasoning; paradigm-level reorganization calls for something closer to Kuhnian gestalt shifts. There is no reason to expect these different tasks to be governed by a single methodology, any more than we would expect carpentry and surgery to be governed by a single manual technique.

The pluralist position is not merely negative — not merely the rejection of a unified theory. It is the positive claim that each type of discovery strategy has its own rational structure, that the structure is intelligible, and that it can be articulated, taught, and evaluated. The absence of a universal methodology does not imply that discovery is irrational or that all discovery strategies are equally good. It implies only that the rationality of discovery is locally constrained rather than globally determined.

9.5 Discovery, Justification, and the Integrated Picture

A final theme worth emphasizing is the integration of discovery and justification. The Reichenbachian separation was motivated by legitimate concerns about psychologism and about the logical autonomy of evidential relations. Those concerns remain valid. But the sharp separation has obscured the ways in which discovery and justification are mutually constraining in actual scientific practice.

The criteria used to evaluate hypotheses in the context of justification — simplicity, explanatory scope, coherence with background theory — are also criteria that guide the generation of hypotheses in the context of discovery. Scientists do not generate hypotheses in a vacuum and then submit them to evaluation; they generate hypotheses that are already shaped by their understanding of what counts as a good explanation. The context of justification constrains the context of discovery, and the success or failure of discovery feeds back into the refinement of justificatory standards.

This integration is particularly visible in the DNA case. Watson and Crick did not generate an unconstrained range of hypotheses and then test them against the data. They worked within constraints — chemical, crystallographic, biological — that sharply limited the space of viable hypotheses. Those constraints derived from the justificatory context: from established chemistry, established X-ray techniques, established biochemistry. The discovery process and the justificatory standards were not temporally separated phases but simultaneous aspects of a unified inquiry.

The integrated picture has normative as well as descriptive significance. If justificatory standards constrain discovery, then improving our justificatory methods — developing better confirmation theories, better experimental designs, better statistical techniques — is also a way of improving the discovery process. Conversely, attending to the strategies that have been productive in discovery — the kinds of analogies that have generated fruitful research programs, the kinds of computational heuristics that have reliably led to empirical laws — can inform the development of justificatory standards.

9.6 Concluding Assessment

Notes compiled for PHIL 358a: Scientific Discovery, Winter 2023. Prof. Doreen Fraser, University of Waterloo.