|
xiii Abductive inference, commonly called inference to the best explanation, is reasoning from given data to a hypothesis that explains the data. Abductive inference is very common in forensic evidence... Abductive inference has been recognized as centrally important in artificial intelligence (AI)... This book presents a clear account of abduction accessible to non-specialists in the philosophy of science or computing.
p.3 A conclusion drawn by abductive inference is an intelligent guess. But it is still a guess, because it is tied to an incomplete body of evidence. As new evidence comes in, the guess could be shown to be wrong.
p.6-7 As well as being important in scientific and legal reasoning, abduction is abundant in everyday argumentation and in everyday goal-directed reasoning of the kind that is currently the subject of so much interest in artificial intelligence. An excellent and highly useful account of the form of abductive inference has been given in the influential work of Josephson and Josephson (1994). Their analysis is quite compatible with the account given by Peirce. They described abduction as equivalent to inference to the best explanation.
p.10 Peirce... stated the very important thesis that "all deduction is merely the application of general rules to particular cases."
p.11 the hypothesis works as a best explanation, given what is known and what is not known in the case.
p.14 Peirce argued that the question of abduction is really the question of pragmatism... his own description of how abduction works... a form of reasoning that has instrumental value in the process by which scientific discovery is made possible through the formation and testing of hypotheses.
p.19-20 A pragmatist, Schiller argued that scientific discovery has its own logic, one different from deductive logic. He saw deductive logic as "static" and argued that there should also be a dynamic logic applicable to cases of arguments in which knowledge is growing (Schiller, 1917, p. 273). However, many philosophers doubted that there could be any such thing as a logic of discovery... A change came with the advent of artificial intelligence. It became apparent to scientists who were engaged in building robots to carry out practical tasks and in creating software that could process information and automate tasks that abductive reasoning is vitally important.
p.22 According to Magnani (2001, p. 19), there are two meanings of the word "abduction." In creative abduction, the task is to generate plausible hypotheses. In a second kind of abduction, called inference to the best explanation, the task is to evaluate the hypotheses.
p.26 Defeasible arguments are ones that can be acceptable at the moment even though in the future they may be open to defeat. New evidence may come in later that defeats the argument. Hence a defeasible argument may be defined as one that is now rationally acceptable even though it may fail to retain this status (Pollock, 1987).
p.28 In a linked argument, both (or all) premises are functionally related to support the conclusion. In a convergent argument, each premise is an independent line of evidence to support the conclusion.
p.29 Abduction also relates to hypotheses that are accepted provisionally, often for practical reasons, or to guide an investigation further along.
p.31 As Peirce's account makes clear, abduction is a kind of supposition-based reasoning that proceeds by the construction of a hypothesis. A hypothesis is a provisional guess that may have to be given up later, when more experimental evidence comes in. So abductive reasoning is presumptive in nature. The burden of proof is not there. A guess is allowed, even if there is very little or no firm evidence to support it yet. But the hypothesis has to be given up if later evidence to the contrary falsifies it.
p.32 What should really be emphasized is that plausible reasoning is based only on appearances, on impressions of a case that could turn out to be misleading once the case has been studied in more depth.
p.32-33 Plausible reasoning is steering an evidence-gathering but open minded dialogue ahead through a mass of uncertainties in a fluid situation by making the presumptive inferences that point to the best path ahead.
p.37 argumentation schemes can be used to fill in missing premises in many common arguments.
p.41 argumentation schemes represent a different standard of rationality from the one represented by deductive and inductive argument forms. This third class of presumptive (or abductive) arguments results only in plausibility... seeming to be true can be misleading. You can go wrong with these kinds of arguments.
p.55-56 Knowledge-based reasoning systems in AI can provide three types of explanations: trace explanations, strategic explanations, and deep explanations (Moulin et at., 2002, pp. 174-76). A trace explanation exhibits the rules and facts used by the system as premises that led to the conclusion put forward... Strategic explanations place an action in context by revealing the problem-solving strategy of the system used to perform a task (Chandrasekaran, 1986). Deep explanations are characterized by a capability to separate the knowledge base of the system from that of the user. In the reconstructive approach to deep explanations (Wick and Thompson, 1992), the knowledge base of the user, rather than that used by the system, is employed in the explanation.
p.71 The thesis that explanations arise from why questions has been advocated in the literature of analytical philosophy from time to time. For example, van Fraassen (1993) postulated the theory that an explanation should be seen as an answer to a question of the form "Why A?" where A represents a fact to be explained. In this sort of approach, an explanation is viewed as a dialogue made up of a question along with the answer to it... As van Fraassen (1980, 126-29) observed, the why question generally does not simply ask for an explanation of a fact, but asks for it in a contrastive way, as in the question, "Why this rather than that?"
p.79 The first point that needs to be understood is that explanation needs to be defined in the dialogue structure as a distinctive type of speech act.
p.82 In the dialogue model of explanation, the text of discourse of the dialogue exchange between the two parties is the basis of evidence enabling an observer or critic to judge how successful the explanation was.
p.82-83 Six basic types of dialog have been recognized as centrally important in argumentation theory. In a persuasion dialogue, one part tries to persuade the other to accept a particular proposition... The inquiry is a type of dialogue in which an investigating group tries to prove some designated proposition, to disprove it, or to show that it cannot be either proved or disproved... The other types of dialogue... are negotiation, information-seeking dialogue, eristic (quarrelsome) dialogue, and deliberation.
p.97 In chapter 1, we saw that AI systems use forward chaining of reasoning from a given database or set of "facts" using a set of rules to generate conclusions derived from the facts. Reasoning is here seen as forward chaining of inferences, and in some instances it could also be a sequence of backward inferences... This model of rationality, which is very common in AI (Prakken and Sartor, 2003, p. 505), is called the inference-based model. But Prakken and Sartor (2003, p. 505) cited another model of rationality they called the procedural view, which embeds a chain or reasoning into a dialogue game, or formal dialogue structure. In this model, a rational argument takes into account "the very possibility of counter-arguments." This procedural notion of rationality, associated with the new field of computational dialectics that is growing in AI, offers some useful tools that can be used in conjunction with the dialogue model of explanation.
p.102 Rationality, in this view, should not be confined just to things you can know or prove beyond doubt. It should be a broad enough concept to include rational opinions that can be supported by good reasons that are, nonetheless, inconclusive (Prakken and Sartor, 2003). In this view, a defeasible argument can be a rational argument if there is enough support behind it in a dialogue to shift the burden of proof or refutation to the critic who opposes it.
p.104 According to this approach, argument is seen as a process (Bench-Capon, 1997, p. 258) made up of a chain of inferences used for some purpose in a dialogue game (Gordon, 1995). The dialogue game has a formal structure in that it allows for certain kinds of moves to be made by either party. These moves typically include the asking of questions and the putting forward of arguments to support a claim. As applied to legal argumentation, the dialogue game can be adversarial (Lodder, 1999).
p.105 Reasoning is a chaining together of steps of inference. An argument is a conversation exchange between two parties in which the two parties reason with each other. So argumentation uses reasoning and is based on reasoning. But reasoning is a context-free notion. It only involves a chaining of propositions.
p.122 In chapter 3, it was shown how reasoning typically takes the form of a forward chaining of defeasible modus ponens inferences that are not deductively valid. It is argued in chapter 4 that many common arguments used in everyday reasoning have the form modus ponens or a form similar to modus ponens but are not deductively valid.
p.129 Presumptive argumentation schemes... involve forms of argument that are neither deductive nor inductive. They represent arguments that are subject to defeat but can be tentatively acceptable to form a hypothesis that enables an investigation to move forward.
p.131 Aristotle defined argumentation from consequences as a distinct form of argument, or what he called a "topic" in Topics... "When two things are very similar to one another and we cannot detect any superiority in one over the other, we must judge from their consequences; for that of which the consequence is a greater good is more worthy of choice, and, if the consequences are evil, that is more worthy of choice which is followed by the lesser evil." ...The general rule suggested by Aristotle's formulation is as follows: all else being equal in a case, choose the action that has the greater preponderance of good consequences as far as one knows from the information given.
p.132-133 This type of argumentation in the positive or negative form, may be defined as argument for accepting the truth (or falsity) of a proposition that recommends a course of action by citing the consequences of accepting (or rejecting) that proposition... This argumentation scheme can be reformulated in such a way that it is seen to be based on an implicit premise that has the form of a generalization... The modus ponens form of argument from consequences brings out how this type of argument is normally used in a typical case of deliberation. Clearly it is a form of practical reasoning. It is a normal, and indeed a very important, kind of argumentation used in deliberation. It tends not to be deductively valid, however.
p.139 The third type of generalization is the plausibilistic generalization of the form "F are generally G, but subject to exceptions." An example is "Birds fly."
p.140 Alfred Sidgwick noted (1893, p. 23) how logic sometimes assumes that only the universal generalizations can "properly serve as ground of inference." But Sidgwick also observed (p. 23): "It is comparatively seldom in actual argument - never, perhaps, where a really difficult or disputed question is raised - that we are able to rest our case in a single faultless generalization..." ...Sidgwick was a precursor both of informal logic and of the recent flowering of artificial intelligence research on defeasible reasoning. He recognized the importance of defeasible generalizations, and the inferences based on them in practical reasoning in everyday argumentation.
p.140 Josephson and Josephson (1994, p. 30) wrote: "The universal quantifier of logic is not the universal quantifier of ordinary life, or even of ordinary scientific thought." In their view, "reasonable generalizations are hedged" (subject to exception). Moreover, they added (p. 23): "Predictions from hedged generalizations are not deductions."
p.142 A weight of plausibility in favor of a proposition's being acceptable only gives a tentative reason for accepting that proposition, subject to doubt and subject to potential retraction.
p.143 an instance of argument from sign. A sign or indicator is observed, and then, because the sign is linked to a certain condition, that condition or thing is concluded to be present... So what good is inference if the conclusion may turn out to be wrong? The function of inference is to make a guess or hypothesis that can lead to testing. Once the tests have been made, the findings may confirm the guess, or they may show it was false. Either way, knowledge is gained about the patient's diagnosis. If the initial guess can be ruled out, then other diagnoses can be explored and tested.
p.144 According to Patel and Groen (1991), the efficacious use of purely forward reasoning is the distinguishing mark of an expert, and backward reasoning is used to tie up "loose ends,"... Generally, then, Patel and Groen found that medical diagnostic reasoning is based on a combination of forward and backward chaining.
p.145 Peirce's analysis, as shown in chapter 1, viewed abduction as a process of forming a hypothesis that can be used as a tentative step in an investigation to explain some observed data... Peirce (1965II, p. 375) saw abduction as a process of inference to the best explanation, "where we find some very curious circumstance, which would be explained by the supposition that it was a case of a certain general rule, and thereupon adopt that supposition." The conditional part of the kind of inference alluded to by Peirce is the use of a "general rule." What is this general rule? Presumably, it is a medical rule of thumb to the effect that, when examining a patient... one possible explanation would be... But of course, all kinds of other explanations are possible... Abductive reasoning is not deductively valid. What is clear... is that it is some type of inference to the best explanation... The abductive inference narrows down the range of possibilities by weeding out the less plausible ones, restricting the search to more plausible ones or even perhaps to one that stands out as highly plausible. But it does not rule out all the other possibilities. It is not deductively valid... it would be a kind of fallacy to portray it in this way.
p.158 Causation is an unsolved problem that affects fields as diverse as science, law, medicine, and history.
p.175-176 One of the most important and convincing applications of abduction is in the kind of reasoning so common in medical diagnosis... this kind of reasoning is typically based on argument from sign. Recent initiatives in using computer systems to aid in this kind of reasoning have been highly successful... The conclusion can be seen as a hypothesis that is justified by the rules and facts. Thus the rules and facts can be said to give reasons to support the hypothesis, or conversely, the hypothesis can be said to offer an explanation of the facts.
p.178 What is important to realize is that the abductive reasoning used in the process of diagnosis is not only a form of reasoning in which premises are accepted and a conclusion proved or supported at the end of the sequence. It is also a dialogue process.
p.180 When producing a hypothesis from a knowledge base composed of a set of facts and rules, a medical expert system can be seen as carrying out a form of causal reasoning. This is because the hypothesis links physiological mechanisms to disease manifestations.
p.181 Peirce saw abduction as the central form of reasoning in the process of scientific discovery. He saw it as representing the creative stage of scientific discovery in which guesses are made by formulating new hypotheses that have not yet been verified or perhaps even tested.
p.181-182 Peirce made room for a discovery stage that precedes the testing and verification or falsification stages of scientific reasoning. This discovery stage is creative, and thus it can be likened to a dialogue process in which questions are asked and a dialogue moves back and forth as different hypotheses are formed and judged as more plausible or less plausible. If this approach is right, the dialogue model can be used to provide a framework to show how abductive argumentation moves forward by formulating a sequence of questions, these guided by the answers and by the incoming data they provide.
p.182 Causal reasoning, in the theory proposed above, is not only abductive but also dynamic. The causal conclusion takes the form of a hypothesis that is derived by inference to the best explanation of the data. But as new evidence comes into the case, the hypothesis can be better and better confirmed, making it more and more plausible as critical questions are answered... This dynamic sequence of reasoning is one of retesting and improving the causal hypothesis as each new bit of evidence comes in. Seeing such a sequential buildup of evidence for a causal hypothesis in a case is just the sort of evidential framework that can and should be used to evaluate argumentation from correlation to cause. So it is just this sort of dialogue framework that allows rational argumentation to deal with cases where the post hoc fallacy is an issue. Of course, such reasoning is not deductive in the most common and controversial kinds of cases. It does not seem to be entirely inductive, either. It seems to be an abductive kind of reasoning that is based on defeasible causal generalizations and is plausibilistic in nature.
p.182-183 Simmons (1992) has represented the generate-test-and-debug (GTD) paradigm as a model for argumentation in a scientific inquiry. The sequence starts with the formulation of a problem and a hypothesis representing a possible solution. The next step is the testing of the hypothesis. The debugger then modifies the hypothesis and resubmits it for testing. Alternatively, of course, the test could falsify the hypothesis, thereby refuting it. Given that kind of outcome, an alternative hypothesis needs to be considered... As the hypothesis is continually retested, experimental support for it builds up. The hypothesis becomes more and more plausible. It is never regarded as beyond testing, however, even thought it may become highly confirmed at an advanced stage of testing
p.185 [premise] (1) My neighbor's roof gets wet whenever mine does... is seen as a generalization that is subject to exceptions. Perl (2000, p.1) drew the general conclusion that "most assertive causal expressions in natural language are subject to exceptions, and those exceptions may cause major difficulties if processed by standard rules of deterministic logic." [JLJ - same thing applies in playing a complex game of strategy. Any heuristic "rule" or algorithm for generating "maybe" moves is likely to fail in certain hard-to-determine-ahead-of-time exceptions.]
p.187 The particular explanation selected as cause, or "best" explanation, is the one that best fits the data... A causal hypothesis is tentative, and said to be relatively plausible or not, comparable with other explanations. It is generally open to defeat, and therefore best regarded as defeasible, in legal argumentation and at the stage of scientific argumentation. As new information comes into a case, the hypothesis that seemed the most plausible one may be replaced by another hypothesis that now becomes the most plausible one, given the expanded database.
p.190 What we see in this case is a forward-backward pattern that is also characteristic of abductive reasoning as a process of inference to the best explanation.
p.203 The abductive theory works best when applied to particular cases where the outcome has actually occurred because, when it comes to predictions, we never really know whether a set of conditions is sufficient.
p.204 The notion of argument chaining is fundamental to both the forward and the backward argumentation schemes for abductive inference... Whether you choose to use the argument diagram or the search graph of the kind used in heuristics, the structure nicely represents the chaining forward and chaining backward of a sequence of reasoning.
p.220 commitment, in the sense of the term appropriate for dialogue theory, is not the same as belief. Commitment is more like acceptance... Commitment... is determined purely by the moves made in a dialogue, where some record has been kept of those moves and where there are rules about whether some move is appropriate or not.
p.225 Schaffner (1980, p. 179) has argued that Hanson and his critics in their writings on abduction never distinguished clearly between two kinds of logic. One is what Schaffner calls a logic of generation, meaning a method for articulating a new hypothesis. The other he calls a logic of preliminary evaluation, meaning a way in which a "hypothesis is assessed for its plausibility." Schaffner's distinction corresponds at least roughly to a distinction long made in logic between two uses or functions of logical reasoning. One is the use of standards of inference to test arguments to see whether they meet standards of structural correctness. This use is highly familiar in logic as conventionally taught. But there is also a longstanding tradition of using logical reasoning to invent arguments called the ars inveniendi, or art of finding. This process looks not for all possible arguments but only for the plausible ones (Kienpointner, 1997, p. 225).
p.228 As Peirce showed, the value of abduction is that it enables a selection of a hypothesis from among a set of alternative possible explanations. The hypothesis can then be tested... What makes abduction useful is that it can lead to the use of more exact forms of reasoning once a plausible hypothesis begins to take shape. It is not the hypothesis itself that is so valuable, but how the hypothesis turns out to be useful in the wider process of investigation as the dialogue moves forward to a conclusion.
p.231 Defeasible arguments are open ended, meaning that they are subject to defeat in the future by a sequence of argumentation that is not known yet. they are not conclusive and closed in the way deductive arguments are, and they are even more open than inductive arguments are.
p.232 This characteristic of defeasible arguments can be formulated as follows... When a proponent has put forward a defeasible argument at some move in a dialogue, the proponent must be open to giving it up and admitting its defeat at any future move should the respondent bring in new evidence that defeats the argument.
p.234 What is characteristic of abductive reasoning is that it is used at the early stages of a dialogue, before epistemic closure has been achieved and very often even before much empirical testing has been carried out. That does not mean that abductive reasoning is chaotic or pure intuitive guesswork. It has certain characteristics as a form of argumentation. And one of these is its defeasibility. It is a type of inference that needs to be seen as used in a context of dialogue that is still open.
p.274 An abductive inference often has small evidential worth by itself, because it is merely a conjecture that may be based on localized evidence. Its primary value is that it can enable an investigation or discussion to move ahead, building up a mass of evidence for one account, as opposed to a contrasting account that is supposed to enable better understanding of what was queried.
|
