Cognitive Science and Folk Psychology: The Right Frame of Mind | Pim Haselager

Chapter 3: The Frame Problem

This chapter focuses on the frame problem from a more cogntive science and mechanical rationality interest.

In general, Pim distinguishes two camps of people discussing the frame problem (FP from now on): The deductivist vs the non-deductivist approach. This distinction more or less follows the "neats" vs "scruffies" in AI.

The deductivists (I) claim:

  • FP is a purely technical problem for logic
  • They are concerned with intelligence in general, not specifically human intelligence.

The non-deductivists (II):

  • FP is a broader problem when trying to understand the psychogical mechanisms of common-sense reasoning.
  • In this broader sense, the problem is more how people reason and act within a changing environment.

Within this second camp, Pim makes two further divisions:

  • (II A) "sees the frame problem as something to be solved by the development of a good representational format" (76)
  • (II B) "expects progress from a better understanding of non-demonstrative inference" (76)

(I) The Deductive Approach to FP

The FP is originally coined by McCarthy and Hayes (1969) in the context of deductive logic.

As said above, they were not specifically concerned with human intelligence.

Intelligence in a GOFAI sense is:

a. Epistemological component: We need some model/representation of the world (in logical language) b. Heuristic component: We then have problem-solving and planning techniques for deductive goal-oriented reasoning.

A "situation" describes the whole universe at a timepoint. A "fluent" is a partial description of a situation. (77)

What we have in practice is a set of sentences about facts.

The FP in this context is essentially what is also called the inertia problem (78):
not only do we need logical rules to indicate what changes as a result of some action, we also need rules to explain what does not change because of that action.

For instance, in order to know that the color of someone's hair has not changed after using a key, the system has to deduce that this fact is (still) the case in the new situation. But it can only deduce this fact if there are rules explicitly stating that using a key does not result in a change of hair color. Since most events or actions leave most facts unchanged, it follows that a great number of rules are needed to specify non-changes (Hayes, 1987, p.125) (77)

These rules for what stays the same are called frame axioms.

In a technical sense, the problem is that you now have a system that spends most of its time deducing things that do not change.

McDermott admits that deductive logic cannot solve the inertia problem. Also see McDermott on 202006211609-tyranny_of_technicality.md

What we need is a rule that condenses "all frame axioms into one general rule stating that if a fact is true in a situation and cannot be proven to be untrue in the situation arising after an event, that it is still true (McDermott, 1987b, p.115)" (78)

However, if you use this rule you have non-monotonic logic, i.e. "on the arrival or production of new information something that formerly could not be proven untrue now can be proven untrue. Hence, truths can turn into falsities. Many common-sense beliefs and inferences are defensible in the sense that their truth values regularily do change on the basis of new information"

So a non-monotonic system can make theorems disappear when new axioms are added.

Pim argues this is still not the most promising result, for two main reasons (79-80)

  1. Non-monotonous logic is a "cognitive wheel" (cf. Dennett)
    • It is not really applicable to psychological phenomena that are interesting and realistic
  2. It neglects the importance of domain-specific knowledge in common-sense reasoning
    • In common-sense reasoning people efficiently draw from what they know (and pick relevant facts)

(II) The Non-Deductive Approach to FP

In everyday reasoning we do not reason about all the ways in which some actions does not change the world, or what all the possibilities are if it goes wrong. Instead, people reasoning under the assumption that "everything is normal".

If we would find a way to specify "everything is normal" we probably have solved the FP.

The sleeping dog strategy

One attempt in logic to meet this idea, is the sleeping dog strategy: only deduce direct consequences of an action, and ignore the rest.

In fact, McDermott (main proponent of this solution) has argued that all working AI problems in fact make this assumption without issues. (Fodor responded is that these systems are by all standards incredibly stupid). In other words: AI is not bothered by the FP and its logical problems because AI is not committed to finding general solutions. He also claims that consequently the people arguing against his viewpoint, sketch unproportionally hard scenarios that would also be difficult for humans to solve. (McDermott wrote an article called "We're framed!")

However, Pim convincingly shows that McDermott's position does not circumvent the FP when formulated in its more broad (and more interesting) sense. AI has in fact not succeeded in modelling common-sense reasoning under normal circumstances (so this addresses McDermott's allegation that all counter arguments assume abnormal situations ).

The FP strikes back: reformulation

This becomes clear if one realizes that, as Haugeland (1987, p.82,84) has noted, the sleeping dog strategy answers the question of how a system can ignore all the irrelevant updating questions by stating, in effect: 'let everything lie, unless there is some positive reason not to'. This raises the question of how cognitive systems obtain these positive reasons. The obvious reply is that people use what they know about the domain at hand in their reasoning about events. The difficult question is how people manage using what they know to efficiently guide their reasoning. McDermott (1987b, p.120) more or less ignores this design question by suggesting that computational systems are able to deal with normal situations and only fail in situations that are so complex or strange that humans would fail as well. (82)

So in short, the inertia problem can be solved, but this does not solve the problem of "finding out which facts do change" (82). And this is actually required to make an intelligent AI!

I conclude, then, that the frame problem can rightfully be interpreted as a problem of dealing with change, and that this is a computationally far from solved problem. (82)

The real difficulty underlying the frame problem is how the relevant pieces of knowledge are found and how they influence one's understanding of the situation, i.e. how they guide the generation of explanatory hypotheses and bear on their acceptability. (83)

Hey, abduction comes around the corner.

The two strategies for non-deductivists

II A: representational approach:

  • find the 'right' ontological structure of the world
  • represent that ontology is a computationally feasible/efficient way

The representational approach is characterized by a strong interest in questions of what kind of information should be regarded as primary and what kind of representation is most suited for representing this information in such a way that the 'right' inferences will follow almost automatically. (83)

II B: inferential problem:

  • given what one knows and perceive, how can one efficiently generate plausible hypotheses to explain what is going on?
  • induction, abduction, IBE.

The inferential approach focuses on the reasoning process itself and investigates how more global characteristics of what one knows can guide the search for plausible inferences. (83)

(II A) The representational approach

Janlert (Haselager, 84):

  • Explicit vs. implicit representations
  • Primitive vs. non-primitive information
  • Primary vs. secondary phenomena
    • secondary: epiphenomena and reducible

Both what needs to be representation and how (efficiently).

There is little hope for solving the FP problem by explicitly representing all primary information.
Instead, the proposal of Janlert is to "find the right metaphysics and represent it intrinsically (Janlert, 1987, p.2, 37-38)." (85)

Intrinsic representation:

"A representation is intrinsic if a representation of a relation has the same inherent constraints as the relation itself (Palmer, 1978, p.271-272)" (85) i.e. the constraints follow from the structure of the representation.

Mind bender:

"The idea is that the representationally inherent constraints are similar to the constraints inherent in what is represented" (85)

Scale model example: a model is a good model if it's structure is similar to what it models. The parallel: "every consequence of an event in the actual world will be precisely matched by the consequences of the small-scale event in the small-scale representational world." (85)

If you reason long these lines, you may argue is that why logic fails to address the FP is that it's explicit structure does not follow the structure of the world (in particular, that it works with a intrinsic/explict binary). There's no very concrete solutions along these lines though. See p. 86.

So the main point is that for an adequate representation: "investigations of the right ontological structure of the world are essential, together with finding a specific intrinsic way of representing ontologically primary information." (87)

Interesting: vs. the declarative style of logic you could oppose the procedural aspects of knowledge.

Dreyfus and Dreyfus suggest that procedural knowledge (know-how) is a more fundamental form of knowledge underlying our common-sense reasoning abilities. The apparent intractability of the frame problem is the result of AI's view of intelligent behavior as arising out of the processing of elements of information about the world (that is, of a symbolic representation of declarative knowledge) (Dreyfus& Dreyfus, 19876, p.97; see also 1986, p.82-90). Dreyfus and Dreyfus (1987, p.105) argue that the frame problem arises precisely because cognitive science has neglected the procedural kind of knowledge that underlies our skills. In their view, a direct recognition of similarity and relevance is an essential ingredient for modeling common-sense intelligence and avoiding the frame problem. (86)

Q: Does that mean that a connectionist model would not be bothered by the frame problem? (86-87)

(II B) The inferential approach

How can we model higher/central cognitive reasoning processes using a symbolic representation of knowledge?

Fodor argues some key ideas are missing for a successful attempt. Pim argues Fodor needs to be taken seriously.

  • Intelligent systems 'look before they leap'
    • He means that cognitive systems project hypotheses and confirm them.
  • Non-deductive inferences add information
  • "The frame problem arises when one addresses the question of how central systems manage to make plausible common-sense inferences. Importantly, there is no fixed delimitation of the information that is relevant to the fixation of a specific belief; everything one knows might bear on the inferences made (Fodor, 1987a, p.140)" (88)
    • This is a point I've seen reappear many times in different contexts, so this is apparently a very important contribution of Fodor. It's called isotropy. See below.

Fodor vs the sleeping dog solution: do not confuse the solution to a problem with simply circumventing it.

Fodor argues that the sleeping dog strategy does not solve the frame problem but merely changes it from an issue concerning belief fixation into a problem about ontology or canonical notation. As the sleeping dog strategy only works when most facts do not change over events, the difficulty therefore becomes how facts are to be individuated. (89)

So in other words, the sleeping dog solutions assumes you have a "canonical representation", in Fodor's words, that remains unchanged through most of the actions. But this is exactly what you cannot guarantee and what is part of the problem: anything could be relevant for any given inference.

Also, the problem has been moved to finding the right representation. But how do you have a representation with the 'right' ontology that excludes notions such as Fodor's "fridgeon" (each particle in the universe that exists at the time Fodor's fridge is open).

This paragraph was a bit unclear to me. I don't see clearly why the fridgeon is brought into play.

I think the main point is that Fodor sees the FP as a problem of non-deductive reasoning because he is very sceptical about the possibility of finding a representation with the right ontology.

SEP-The Metaphysics of Inertia:

An additional dimension to the frame problem is uncovered in (Fodor 1987), where the metaphysical justification for the common sense law of inertia is challenged. […] As stated above, solutions to the logical frame problem developed by AI researchers typically appeal to some version of the common sense law of inertia, according to which properties of a situation are assumed by default not to change as the result of an action [JK–the ‘sleeping dog strategy’]. This assumption is supposedly justified by the very observation that gave rise to the logical frame problem in the first place, namely that most things don’t change when an action is performed or an event occurs. According to Fodor, this metaphysical justification is unwarranted. To begin with, some actions change many, many things. Those who affirm that painting an object has little or no effect on most properties of most of the objects in the room are likely to concede that detonating a bomb actually does affect most of those properties. But a deeper difficulty presents itself when we ask what is meant by “most properties”. What predicates should be included in our ontology for any of these claims about “most properties” to fall out? To sharpen the point, Fodor introduces the concept of a “fridgeon”. Any particle is defined as a fridgeon at a given time if and only if Fodor’s fridge is switched on at that time. Now, it seems, the simple act of turning Fodor’s fridge on or off brings about an astronomical number of incidental changes. In a universe that can include fridgeons, can it really be the case that most actions leave most things unchanged? […]The common sense law of inertia is only justified in the context of the right ontology, the right choice of objects and predicates. But what is the right ontology to make the common sense law of inertia work? Clearly, fridgeons and the like are to be excluded. But what metaphysical principle underpins such a decision?

There has been criticism that Fodor doesn't know what the FP is. Pim's defense is that for Fodor the FP is a symptom of the more general problem of cognitive science to model central cognitive processes.

If cognitive science is interested in studying the nature of rationality and reasoning, the frame problem is not just a problem, but rather a valuable problem precisely because it directs attention towards more general concerns such as the holistic nature of general rationality. In this sense, the frame problem is, as Dennett (1987a, p.43) puts it, "the whole pudding". (90)

AI and cognitive science:

philosophy tries to see how AI relates and contributes to the study of cognition. Treating the frame problem as a technical issue is missing a chance to deepen our understanding o these more general matters. (90)

To understand the type of non-demonstrative reasoning that takes place in common-sense reasoning, it is useful to study similar and more well documented types of scientific reasoning.

Fodor distinguishes two characteristics of the fixation of belief in science:

Isotropy and Being Quineian

Isotropy:

every established fact can be of importance to the evaluation of a theory of hypothesis. Everything you know may help determine what else you ought to believe. Isotropy stresses the potential relevance of each individual belief and is closely related to the importance of analogical reasoning (Fodor, 1983, p.105-106) (91)

Being Quineian:

the degree of confirmation of a hypothesis depends on characteristics of the whole belief system in which it is embedded. The notions of simplicity, conservatism and plausibility are often mentioned in this respect as important but poorly understood criteria for evaluating a given body of belief (Fodor, 1983, p.107-108). The 'Quineian' feature emphasizes the global character of justification of our knowledge. (91)

How then is FP an inference problem? Example:

"Witnessing an event, we effortlessly generate plausible interpretations of what is going on and what the likely results of the event will be. Solving the frame problem requires insight in how the global characteristics of the knowledge we possess guide our non-demonstrative inferential practices." (91)

The severity of the FP

Fodor argues the FP is a philosophical problem that requires a breakthrough in inductive logic. The problem is that global factors play a role in the fixation of belief. There is no solution to the FP for the same reason there is no good philosophy of scientific confirmation. (92) Fodor frames this as a problem for the psychology of central cognitive processes.

Putnam also thinks the whole AI endeavour is not that impressive yet because it does not solve the problem of inductive inference that is its core problem, according to him.

The bottom line is that Fodor does not argue that cognitive science etc is doomed, but rather that there is virtually no progress because fundamental ideas are missing. (93)

Descriptive and normative approach to non-demonstrative inference

  • psychological-descriptive: how do people actually reason and arrive at conclusion?
  • philosophical-normative: how should they reason if the conclusion is to be justified?

Historically, these have been separated and epistemology has been dominantly normative.

Quine coined the idea of 'naturalistic psychology' where psychological theory is allowed to inform normative theory.

Hume's account of induction:

  • "as a logical a priori principle it fails since its denial does not result in a contradiction"
  • "as an empirical a posteriori principle it begs the question: the fact that in the past everything has stayed more or less the same is only relevant to the future if the principle at issue is assumed to be correct." (94)

Hume thus argues that on a logical level these inferences cannot be justified, whereas psychologically people still make them out of psychological habit.

Fodor's point about key ideas missing relates these two levels in a different way: we cannot accurately psychologically describe central cognitive processes, because we miss the right theory of induction confirmation.

Pim argues that this position is not that accurate: descriptive theories aim at psychological plausibility rather than logical formal correctness. Paradoxes (Raven, Goodman's grue riddle) show that what we consider to be relevant in inferences can in fact diverge from what is logically proven to be relevant (according to a theory of confirmation etc.).

Chapter 4: The problem of non-demonstrative inference

Question that is picked up here: "is it possible to work towards a solution of the problem of non-demonstrative inference without being guided by developments in logic" (98)?

This question is clearly asked in discussion with Fodor, who holds that we have to wait for a breakthrough in inductive (confirmation) logic to solve non-demonstrative inference and the FP and such (basically, being able to express logically which facts have a confirmatory influence on some explanandum).

As a case study, this chapter discusses the work of Thagard which offers a descriptive account of non-demonstrative inference.

Terminology

Types of non-demonstrative inference:

  • Enumerative induction
  • Hypothetical induction (Harman, 1992, p.200-201)
    • "one reasons that something must be (or have been) the case as that would explain an encountered phenomenon."

Schemas:

  • Deduction: p -> q, p : q.
  • Enumerative induction: p, q : p -> q
  • Hypothetical induction: ass. p -> q, observe q, conclude p. (a logical fallacy; affirmation of the consequent)

Hypothetical induction can be further distinguished in abduction and inference to the best explanation. These concepts are often used differently: Pim distinguishes the latter two only for specific aspects of hypothetical induction.

Hypothetical induction is usually called abduction in AI.

Abductive reasoning:

  1. 'abductive rule generation'
    • Thagard focuses on this; claims to follow Pierce
  2. 'abductive rule inference'

Pim reserves the word 'abduction' for the generation of explanatory hypotheses.

Pim reserves the word 'inference to the best explanation' where the evaluation of hypotheses is central (so second aspect); e.g. you infer the one that is most likely, which implies the evaluation of the hypotheses.

N.B. Fodor often speaks of induction as a name for all non-demonstrative inference, which is a bit sloppy. When Pim really means enumerative induction, he says so.

Quite obviously, the two aspects of hypothesis induction are more relevant for the FP than enumerative induction.

These two "epistemic filters" (term by Peter Lipton) both have their own problems.

  1. There must be some mechanism that deals with isotropy when generating and evaluation some hypothesis, i.e. isotropy.
  2. and a mechanism "to determine how holistic characteristics of relevant information in relation to a hypothesis influence the latter's acceptability, i.e. Being Quineian." (102)

Thagard is optistic where Fodor is pessimistic, because he has high hopes for a pragmatic approach: "there is no general, abstract, logical way of dealing with all kinds of non-demonstrative inferences normally performed by cognitive systems. Instead, specific ways of dealing with specific problem situations are required, explicitly taking into account the goals and the context of the system performing the inferences, as well as the influences exerted by the way in which the knowledge is organized in the system (Holland, et al., 1986, p.7; Thagard, 1988, p.39)"

Constraints on approaches to the problem of non-demonstrative inference

See p.103:

  1. stays within the framework of classifical cognitive science
    • symbolic representation system
    • compositional syntax and semantics
    • truth related interpretation of knowledge
  2. explains how plausible and relevant hypotheses can be generated, rather than a huge set of random ones.
    • Cf. the paper on analogy we discussed in TFCA; perhaps fails on this point.
    • "avoid the generation of awkward hypotheses" (103)
  3. accept phenomena that intuitively count as confirmatory evidence for a hypothesis, at the same time rejecting unintuitive ones even if they are logically equivalent
    • Cf. Hempel's paradox
  4. have a grip on holistic characteristics of a body of knowledge important for selecting hypotheses
    • e.g. simplicity, coherence
  5. it does not fix 2,3,4 by defining an a priori knowledge structure.

Thagard's models

Thagard has several models where activations spread through conceptual hierarchies, where concepts are represented as frame-like structures with slots to fill in. He uses these models to computationally model how aspects of hypothetical inference take place, e.g. generating hypotheses.

Generating hypotheses

Generating hypotheses: the main question here is how do you avoid irrelevant or awkward hypotheses?

In this context, Goodman's new riddle of induction is cited.

The thesis that all emeralds are green is confirmed by all green emeralds found. Yet these same instances of green emeralds can also be taken to confirm the thesis that all emeralds are grue, where the predicate grue is defined as: x is grue when it is green before time t (say the year 2000) and x is blue after that time t. The problem is to explain why the first hypothesis should be taken seriously to be confirmed by instances of green emeralds, while the second is not. Alternatively, why do we generate a hypothesis like 'all emeralds are green' on encountering an instance of it, instead of the hypothesis that 'all emeralds are grue'? (107)

Goodman's own suggestion is that serious hypotheses use "entrenched" concepts.

Thagard's solution: "the mechanism of directed spreading of activation avoids the investigation of numerous unnecessary hypotheses. Because inference is primarily studied in the context of problem solving, the concepts involved in the specification of the problem and the goals can start the spreading of activation." (107)

So the idea here is that 1) initially only concepts present in the problem statement will be activated, and 2) that the activation flow is directed by the structure of the knowledge representation. See an example on p. 110.

To avoid putting in all knowledge a priori, such a network must be able to adjust its own conceptual hierarchy. The combination of concepts into new concepts is called coalescence. (112)

This approach, Thagard feels, is capable of dealing with Goodman's 'grue' puzzle. The concept 'grue' is not available and will not be generated unless it is needed by the cognitive system and turns out to be of value. The entrenchment of a concept of proposition is dependent on its having a place in the conceptual hierarchy, its utility in processing and its coherence with other parts of knowledge., and all these factors can be made computationally precise (Holland, et al., 1986, p.235; Thagard, 1988p.36;1992,p.94) (112)

Evaluating hypotheses

Here the relevant paradox is the Raven paradox: "There must be a way to exclude certain kinds of evidence as irrelevant to a specific hypothesis" (114)

Again, in Thagard's model what does not get activated will not play a role for evaluating a hypotheses, again underlying the importance of the structure of the knowledge representation.

But the crucial difference is that relevance is here determined by structural relations in the concept hierarchy, not logical ones.

As a result, such a computational model may capture how people in fact reason and what they deem is relevant, rather than what they ought to think based on logical equivalences. This is relevant for the Raven paradox, where we intuitively see black ravens as a confirmation of the statement 'all ravens are black' whereas we do not we non-black non-ravens as confirmatory evidence.

Descriptive vs normative approach

From the previous sentence it becomes clear that Thagard's approach is "a descriptive-psychological one: how do human beings in actual practice go about making sense of events in their surroundings by generating a plausible hypothesis on the basis of incoming information and their knowledge." (119)

This should be clearly distinguished from the quest of logical confirmation (Fodor's pessimism & hope).

"AI takes schemas to be relevant to the frame problem not because they provide a solution to the problems in the logic of confirmation but because they are important computational aids in representing information effectively." (123)

Isotropy, Being Quineian and concept directed activation models

If you want to express that any fact could be relevant as the conditional in an explanatory hypothesis (isotropy) - and when we depends so much on the structure of these activation networks - then "this implies that all the conceptual structures that could be of explanatory relevance have to be linked to the conceptual structure representing the explanandum." (121)

This ins unfeasible. This in turn requires you to already know, a priori, which facts may be relevant and which facts are not, and this circumvents the issue it aims to solve.

A first objection of Fodor is that much research in AI of this kind evades the real problem by artificially restricting the models, for instance, by not supplying them with knowledge that is deemed beforehand to be irrelevant to the kind of tasks to be confronted by the model. (122)

Fodor:

You don't build the model so that it begins to think about protons when its worrying about tigers. You don't even put any information about protons in, but people have information about protons and they know how not to use it. The question is how do they know how not to use it. (Fodor in Pylyshyn & Demopoulos, 1986, p.167) (122)

"Fodor argues, structural relations between complex representational structures are insufficient to capture the true nature of our reasoning processes. The links connecting chunks of knowledge are mere associations." (123)

Fodor: Reasoning =! association.

If inference relies heavily on association then "You would only be able to understand a world which does what you expect it to." (Fodor in Pylyshyn & Demopoulos, 1986, p.136-137) This was also my thought for the purely analogical account of abduction.

However, the (more modest) claim of "scruffy" AI is that associations can guide inference. (124)

Very interesting section Quine-Dunhem thesis .

We cannot decide a priori on what is a relevant representation structure:

The guidance provided for by links that associate bits of information will be of little value because there is no general way of indicating what is relevant to what. On the basis of the Quine-Duhem thesis, Fodor argues that one cannot once and for all localize the class of data relevant to the fixation of a belief or the confirmation of a scientific hypothesis (Fodor, 1983, p.108-109). Quine argued that the first dogma of empiricism, i.e. that there is a distinction between analytic and synthetic statements, has to be rejected since there is no non-circular way to flesh out the notion of analyticity. Because of that the second, reductionist dogma fails as well; the confirmatory instances of a statement cannot be known a priori (Quiene, 1951/1980, p.40-41). The consequences of assuming that there is no analytic-synthetic distinction is confirmational holism. Our knowledge of 'what confirms what' is a posteriori and contingent (Fodor & Lepore, 1992, p.37-38). Therefore, the endeavour to make a priori or general decisions on structure, on what is relevant to what, is an impossible one. (124)

Yet, another formululation of the FP in AI: "it's just a problem of confirmation that emerges when you try to do a confirmational logic for what's in somebody's head. (Fodor in Pylyshyn & Demopoulos, 1986, p.163)" (125)

Pim reiterates we must distinguish Fodor's concerns as normative; and his problems as well mainly for a normative and logical account. A psychological-descriptive effort is really not that bothered by logical problems! And in practice, psychology/AI does not have to be guided by logic!

Great phrasing!:

The importance of the distinction between these two ways of looking at non-demonstrative inference becomes clear when one considers that from a logical-normative point of view the issue is that everything one knows influences what one ought to believe. But psychological-descriptively, the concerns is about how part of what one kn0ows influences what one ends up believing. (127)

And whereas Fodor concerns the reasoning of an "reasonably idealized scientists", this should in fact not be the point of departure for psychological theory.

Using an 'idealized scientists' as the guiding standard for psychological theory development would be, in my view, a crucial mistake. The point is that an idealized scientists is not a psychologically interesting phenomenon as far as investigations into common-sense reasoning are concerned, but rather an epistemological one. (130)

We should not let normative logic restrict descriptive work. Especially because "cognitive science studies 'satisficers', not 'optimizers'." (130)

Idealized scientists are not the proper subject matter for psychology. An examination of these abstract phenomena may surely yield interesting information for psychology, for instance, by indicating points where psychological processes take a differenct course than a purely logical analysis would allow or where psychological processes typically take a specific route which from a logical point of view remains inexplicable. But precisely for this reason, they do not provide adequate benchmarks for cognitive science investigations of psychologically plausible reasoning. (132)