This page contains abstracts for a number of papers that I have authored or co-authored. Unfortunately, this page is very much out of date. I hope to find time to update it soon. My resume on this site is much more up to date and should be consulted for publication information. Please email me for information about reprints. Some of the papers have been published or presented, others are in various stages of preparation. Click on a title in the list immediately below to view the abstract. Unlinked titles currently have no abstracts (usually because the paper itself is not yet completed). Whenever possible, links are provided from the abstract to viewable or downloadable copies of the entire paper in various formats as well as to other materials and WebSites. Where indicated, hardcopy versions are available via snailmail. Just email your request to me at steve_kemp@unc.edu. Overall, there are five categories of papers: Situated Learning, Neural Networks, Induction and Abduction (i.e., ampliative logic and reasoning), Language and Verbal Behavior, Psychometrics, and miscellaneous papers fitting in none of the above categories.
This paper is a chapter published in an edited book on The Structure of Learning Processes, edited by Jaan Valsiner and H-G. Voss. It is an effort to sketch out solutions to problems inherent in working out a radical behaviorist perspective within the framework of action theory. Early drafts were halved twice, so the result is unfortunately very dense. (That's the comment of an overly generous friend.) In addition, the paper is over five years old and is very prospective and abstract compared to my more recent work on this topic. However, it did serve to put my position on record and connect up some of my more important influences, especially Vicki Lee, of whose version of radical behaviorism I am, then and now, an uncompromising advocate and devoted follower. The notion of the "predicate object distinction," however, has been abandoned in more recent work for other formulations.
Abstract: This chapter constitutes a proactive effort to build a foundation for animal learning theories based upon an alternative to stimulus-response psychology, the psychology of action. It consists of three sections: First, what sorts of structure are made evident by a radical behaviorist view of action? Second, an argument that the action theory perspective brings into question the usual choice of the basic unit of analysis in animal learning theories and suggests an alternative. Finally, a typical example of operant learning phenomena, operant conditioning of the pigeon, is redescribed using the action-theoretic language. Lee's (1988) radical behaviorist action theory is contrasted and compared to other theories of action. Important features of her views, including a dismissal of the older stimulus-response terminology, a commitment to the inherent means-ends structure of action, a focus on the distinction between action (that is, accomplishment) and bodily movement, and a central position for the commonsense notion of action as things done by the organism, rather than merely reactions to stimuli place Lee squarely within the action theoretic camp. Lee's focus on the individuation and explanation of action in terms of its consequences, rather than its antecedents, are in distinct contrast to other action theories and reflect her commitment to radical behaviorism. The present paper encapsulates this notion of psychological explanation in terms of means made up of bodily movements and ends understood as further actions. We perform deeds in order to gain the opportunity for further action. By classifying indidual acts into action-classes linked to general goals, psychological explanation is grounded in the consequent effects of action. Next, different logical relations in both the macrostructure and the microstructure of conduct are examined. An argument is offered to the effect that we can distinguish what Austin (1960) called basic actions by means of a linguistic distinction called the predicate-object distinction. Finally, operant conditioning of the pigeon is described in terms of re-orientations of the head and pecks. In Mechner's (1992) terms, the keypeck is an operant with the peck as its terminating sub-operant. Simply put: every keypeck is a peck, but not every peck is a keypeck.
Availability: Click here to view online version. Valsiner and Voss' book is in print and may also be at your local University library. Regrettably, the publishers have a bizarre policy of not selling reprints of the chapters. (Apparently, my money isn't good enough for them.) A viewable and downloadable version is available in html. Downloadable versions are also available in PostScript and RTF. (Warning: the RTF version is not yet tested.) If none of these will serve, a small number of hardcopy reprints can be had via snailmail.
This paper is a reprint of an article co-authored with Dave Eckerman and published in The Mexican Journal of Behavior Analysis in their special issue on the nature of environmental contingencies. The paper foreshadows the development of the In-Situ simulation testbed, whose first use is reported in the paper on Situational Descriptions abstracted below. The most important feature of the paper, given that the system described in prospectus is now in operation, is the discussion of how computer simulations serve a distinct explanatory purpose from most other types of psychological theories. Although the content revolves around the philosophy of science and the theory of explanation, here again, the target audience is behavior analysts.
Abstract: It is useful to "model" both contingencies and behavior. A reinforcement schedule arranged in the laboratory is a working model of a real-world contingency in the sense that a model of an airplane is a model. It simplifies a natural contingency, but it really will "fly." Thus, a reinforcement schedule is both a model of a contingency and it also IS a contingency, albeit a synthetic one. The authors argue for the use of working models of organisms, computer programs that take in models of stimuli as input and produce models of responses as output, all on a real-time basis. Computer simulations of working models, called in situ simulations, are more complex than molecular simulations such as Shimp's Associative Learner in that spatio-temporal patterning, not just temporal patterning, is modeled. Working models of organisms, it is claimed, will allow for testing of behavioral theories of many types without the difficulties attendant upon theorizing that concerned Skinner (1950). Advantages to using working models include: (1) Analysis of behavioral theories at levels of observation more microscopic than those readily obtainable elsewhere. (2) Easy integration with neural network models. (3) The possibility of statistical testing in conjunction with the experimental analysis of behavior.
Availability: This journal is relatively hard to get ahold of, though I recommend it for your library. The original article appeared both in English and translated into Spanish. Hardcopy reprints are available via snailmail (English language only). Online versions in html, PostScript, and RTF are scheduled for the near future. For anyone who desparately needs a Spanish-language reprint and has no access via inter-library loan, special arrangements for copies of the Spanish language version may be possible. The full reference is:
Kemp, S. M. & Eckerman, D. A. (1995). Direct analysis of contingencies using working models. Revista Mexicana de Analisis de la Conducta. (The Mexican Journal of Behavior Analysis), 21, 27-46.
This paper is a reprint of a talk co-authored with Dave Eckerman and presented at the 1995 convention of the Association for Behavior Analysis at a symposium sponsored by the The Mexican Journal of Behavior Analysis. It post-dates the article in that journal by the same name (abstracted above). The major difference between this talk and the journal article above is the inclusion of a discussion on the relation between Dretske's (1981) version of information theory and the re-analysis of the operant conditioning of the pigeon. This analysis eventually led to the current analysis in terms of Situativity theory. (Barwise & Perry's, 1983, situational semantics also depends upon Dretske's theory.)
Abstract: Contingencies are facts of the world surrounding the organism. We can describe these facts with if... then... rules. What allows subject organisms to be controlled by the contingent facts of the world, given only the flux of sensory signals at the nervous periphery? Topographies are necessary components of behavior. A robot, like a natural organism, must produce behavior with some topography. Therefore, we propose the Robot Test. Program any behavioral theory into a (very, very simple) robot. If the robot can run on that alone, the theory is complete. We program a simulated animal body inside a simulated environment inside a computer. The entire system is called an in situ simulation. In information-theoretic terms, a key-peck involves the transmission of information from the pigeon to the operant chamber. The source is the pigeon's beak. The receiver is the key switch. At the source, the information is the peck, the thrust of the pigeon's head, that is, the topography of the response. At the receiver, the information is the position of the key switch. At the source, there is an enormous amount of information in analog form, the topography. At the receiver, there is exactly one bit of information, the function: the circuit is either closed (bit = one) or open (bit = zero). That one bit of information is the message. Any other information, that is, the topography, is just channel. Likewise, in the case of simple discrimination, it is the message that controls behavior. For both stimulus and response, what behavior analysts call function is, in information theoretic terms, the message. What behavior analysts call topography is the channel. (This notion of a stimulus topography is very different from the so-called "stimulus control topography." Here, topography is what does NOT control behavior.) Direct analysis means that the contingencies are provided to the simulation embedded in a simulated sensory flux. In the usual sorts of theoretical behavior analysis, the theoretician describes the relations between contingency-rules and operants. In essence, the theoretician simplifies the problem by extracting the contingency-rules from the contingencies before describing those functional relations. In direct analysis, as in the experimental analysis of behavior (EAB), the contingencies arrive as is. The theory must do its own extracting, just as the organism does.
Availability: Hardcopy reprints are available via snailmail. Online versions in html, PostScript, and RTF are scheduled for the near future.
This paper is a preprint of an article co-authored with Dave Eckerman submitted to The Journal for the Experimental Analysis of Behavior on situated learning. It has been returned for revision and resubmission. The point of the paper is to illustrate how a single standard for evaluating and testing neural network models can be constructed using situativity principles. The data used is from the operant psychology laboratory, so there's a lot of insider baseball meant for a behavior analytic audience. Nonethelesss, this paper represents the current state of our research on this topic. And upcoming papers (abstracted below) will be tying these analyses to various other approaches.
Abstract: We present a new system designed to make predictions for computational theories of learning such as artificial neural networks. Predictions are generated by the system at the level of detail of the behavioral protocol and are displayed as cumulative records. The cumulative record indicates what responses would be emitted by a real organism whose nervous system operated according to the theory modeled by the neural network. An extension of Mechner's (1959) notational system for the description of behavioral procedures provides the basis for the new system. The original notation is extended to take account of new developments in situational semantics. The extensions distinguish situations (environmental circumstances) from stimuli (information about situations available to the organism) and distinguish deeds (bodily movements) from responses (situational changes due to actions). Predictive techniques are illustrated by evaluating a simple theory of operant learning with respect to three behavioral procedures: continuous reinforcement, fixed ratio reinforcement, and fixed interval reinforcement.
Availability: Click here to download.
This paper has been researched and is currently being written. A target journal has not yet been selected. The idea is to relate Guthrie's historical concern with actions and bodily movements to the contemporary re-emergence of commitment to biologically-based models of learning such as neural networks. The paper is targeted for a general psychology audience and should be of interest to anyone interested in action theory, neural networks, or the history of psychology.
Abstract: Guthrie's (1940) distinction between movement and act is reviewed. His original motivation for this distinction is considered. The historical reason why neobehaviorists neglected this distinction is examined. It is argued that, with the emergence of biobehavioral models in behavior analysis, the issues that originally justified Guthrie's concern with the movement/act distinction have re-emerged. The relation between the movement/act distinction and the possibility of so-called type-type reduction of behavioral events to biological events is discussed. Three approaches to dealing with the relation between movements and accomplishments are considered: Guthrie's (1933) correlational approach, the cognitivist representational approach, and Lee's (1983; 1988) action-theoretic approach.
Availability: This paper is in preparation. When drafts are available, the announcement will be made here.
This paper has been researched and is currently being written. A target journal has not yet been selected. The idea is to raise issues surrounding an attempt to provide operational definitions for two widely used but seldom discussed terms in behavior analysis: topography and function. The paper is narrowly targeted for a behavior analytic audience but should be of interest to anyone interested in the epistemic foundations of psychology.
Abstract: Despite wide usage within behavior analysis, an operational definition of the distinction between behavioral function and behavioral topography is not readily obtained. Issues surrounding this distinction will be examined with an eye to providing such a definition. Ambiguities in the usage of the two terms are highlighted. Weiss' (1924) claim as to the logical incommensurability of topography and function is presented as the central difficulty in establishing an operational definition. Weiss' terminology is expanded to account for Skinnerian analyses. The relation between identifying behavioral function and constructing psychological explanations is clarified in terms of (a) the distinction between empirical law and theory, (b) Skinner's identification of the proper referents of explanatory terms in psychology, (c) our previously unpublished information-theoretic model of the topography/function distinction. Lee's (1983; 1988) proposed action theoretic solution is re-examined in the light of this analysis. In concert with the abandonment of stimulus-response terminology for the means-ends terminology of action theory, the two meanings of the term "topography" can be related to one another. The movement/accomplishment distinction central to Weiss' analysis is modeled as a special case of Lee's means-ends formulation. The interest-relative component of the topography/function distinction is captured by a hierarchy of means and ends where movement is the lowest level. Finally, a mathematical model of this hierarchy is proposed in terms of Partially Informative Markov Decision Problems (PIMDPs).
Availability: This paper is in preparation. When drafts are available, the announcement will be made here.
This paper is being researched. A target journal has not yet been selected. The idea is to demonstrate how the tools being developed in the recent efforts within the field of Reinforcement Learning to coordinate the tools from Dynamic Programming with the goals of Artificial Intelligence can be used to analyze Reinforcement Schedules. The paper is broadly targeted for an artificial intelligence audience.
Abstract: Recent efforts by Singh and others to formalize the tasks faced by Machine Learning programs as Partially Observable Markov Decision Problems (POMDPs) provide a key to unravelling dynamic components of traditional animal learning tasks that have heretofore defied formal analysis. Reinforcement schedules, an important technique in the study of reinforcement in the animal laboratory, are rarely modeled in Reinforcement Learning simulations. The present paper offers a series of proofs that a wide variety of reinforcement schedules can be modeled as POMDPs. The implications are that Reinforcement Learning models can be tested against models of reinforcement schedules and the simulation results compared to the mass of laboratory data where animals were confronted with those same tasks.
Availability: This paper is in preparation. When drafts are available, the announcement will be made here.
This paper is being researched. A target journal has not yet been selected. The plan is to tie criteria used to evaluate algorithms in Dynamic Programming to the development of a minimalist criterion for computational theories of learning. The paper is targeted for both mathematical psychology and artificial intelligence audiences.
Abstract: Given the broad applicability of Partially Observable Markov Decision Problems (POMDPs) as models of real-world learning situations, as well as the applicability to highly simplified models of basic laboratory procedures under which the performance of animals is invariant and well-understood, it is possible to propose a minimum standard as to what constitutes a computational theory of learning. Performance on a POMDP using a fixed strategy can be modeled a Markov process. Under specifiable conditions, the average reward for a given strategy approaches a limit. All such strategies can be compared for any such task. If a learning theory is modeled as moving through a space of changing strategies, improvement due to learning can be measured as increase reward due to the change to a better strategy. Given a suitable choice of task, a minimalist criterion for learning would be that, starting from any arbitrarily chosen initial strategy, a system governed by a learning algorithm will converge to a set of superior strategies. Examples are provided.
Availability: This paper is in preparation. When drafts are available, the announcement will be made here.
This poster was presented last at the annual meeting of the Association for Behavior Analysis in San Francisco in 1996. It is a first demonstration that a neural network based upon the biology of In-Vitro Reinforcement, rather than the usual basis of Long-Term Potentiation, can perform a Machine Learning task.
Abstract: A neural network simulation is used to generate a series of black and white images. No antecedent stimuli are supplied. The neural network is reinforced according to a percentile reinforcement schedule (p=.30, m=20) with respect to how close the emitted images are to a target image drawn by the user. Gradually, the distribution of images moves closer to the target until the neural network generates an exact copy of the target. The learning algorithm is based on In-Vitro Reinforcement (IVR), a mechanism whereby spontaneous burst rates of CA1 pyramidal cells of the hippocampus can be increased due to dopamine infusion immediately following a burst. Each unit in the neural net corresponds to one pyramidal cell. Each signal output from a unit corresponds to one burst. Independent thresholds for each unit determine the probability of output from that unit. Results indicate that an eight by eight image can be duplicated in anywhere from 25 thousand to 250 thousand cycles, equivalent to a range of approximately 20 minutes to 3 hours 20 minutes in real time. Modifying thresholds using a linear (truncated) learning rule generates a typical asymptotic curve in various aggregate measures of approach to the target.
Availability: As a poster, no reprints are available. However, the results reported here are available online in the paper, Activationist Learning in Selectionist Neural Networks, abstracted below.
This paper was presented at the annual meeting of the North Carolina Cognition Group in Greensboro, NC in 1996. The data is the same as that of the poster abstracted immediately above. The perspective is cognitivist rather than behaviorist.
Abstract: Many cognitive tasks, particularly complex cognitive tasks, found both in life and laboratory are "proactive" in the simple sense that the behavior has no obvious causal antecedents. In animal learning, experiments without explicit antecedents use what are called "free-operant" procedures. One such procedure, known as "shaping," can be used to produce the sort of complex performances said to manifest cognition in humans and, for those who study animal cognition, in animals. This talk introduces a neural network simulation whose learning algorithm is based on a different sort of neurophysiological mechanism, called In-Vitro Reinforcement (IVR), than that used by adaptive neural networks thus far. The behavior chosen for the simulation is a search task, that of duplicating a binary sequence by successive approximations with one bit feedback. The neural network generates a sequence of guesses. The training routine provides binary (yes-no) feedback to shape the network's performance. Eventually, the network outputs the target string.
Availability: Reprints of this talk have not generally been made available. A limited number of reprints can be made available if special circumstances require it. In general, the results reported here are available online in the paper, Activationist Learning in Selectionist Neural Networks, abstracted below.
This paper was published in the March, 1997 issue of The Journal for the Experimental Analysis of Behavior. It is a commentary on a major article by Donahoe, Palmer, & Burgos published in that same issue. The principal point is to challenge those authors presentation of neural networks as reflecting a unified view of learning.
Abstract: In their target article, Donahoe, Palmer, and Burgos (1996) argue that their behavioral approach is response-stimulus (R-S), while their computational approach is S-O-R. An important alternative to this strategy is available: Activationist neural networks, unlike connectionist neural networks, model reinforcement as the strengthening of responding rather than as the strengthening of environmental control of responding. Two activationist neural networks, Clavier and Piano-Forte, are under development. The present commentary attempts to illustrate that, contrary to the target article, selectionist neural networks constructed from elementary units intended as analogues of Skinner's "behavioral atoms" (Stein, 1995) can also offer a unified theoretical treatment of various sorts of conditioning. Commitment to networks "constructed from elementary connections intended as analogues of stimulus-response relations" (Shull, 1995, p.354) is not necessary.
Availability: The March, 1997, issue of The Journal of the Experimental Analysis of Behavior did a special article, with multiple commentaries and an authors' reply, on the utility of neural networks as theoretical tools in the study of animal learning. The article, complete with commentaries (including my commentary abstracted here above) and reply, can be downloaded from the journal's WebSite. JEAB uploads one article per issue to its WebSite. The file is in .pdf format, which requires Adobe Acrobat for browsing and/or printing. Happily, the journal provides a link to the Adobe site from where the Adobe Acrobat Free Reader can be downloaded free of charge. Hardcopy reprints of the single commentary abstracted here are available on request.
This paper is a online technical report written especially for this WebSite. It is intended both to summarize current research on IVR-based neural networks and to document future plans. An extensive online glossary on neural networks is also provided. The effort is to design a document that is accessible not only to specialists in various disciplines, but also to interested lay persons. Comments and feedback are welcomed.
Abstract: The first example of a novel type of neural network with a learning algorithm based upon the neurophysiological research of In-Vitro Reinforcement is presented and described in detail. Various tasks already demonstrated by the network or proposed for future research are described. Three architectures are offered: a basic model called Clavier capable of being shaped to search spaces of binary strings, a smoothed version called Orchestrion capable of searching spaces of integer vectors, and a damped version called Piano-Forte capable of pattern recognition. Details are offered for all three models.
Availability: Click here to view online version. A viewable and downloadable version is available in html. A downloadable version of Appendix C, which provides mathematical details of the algorithms for Clavier, is also available in PostScript and an RTF version is being tested. No hardcopy versions are planned.
This paper was written a few years ago and is awaiting revision. It gives a theoretical account of the conjunction fallacy along the lines of other recent theories of the overconfidence effect and base-rate fallacies. The paper is targeted for researchers in the area of cognitive heuristics and biases.
Abstract: To date, there is no positive, descriptive account of how subjects reason in order to conclude that a conjunction may be more likely than one of its conjuncts. The conjunction fallacy is an empirical effect than seems to entail fallacious reasoning on the part of subjects, but the nature and course of that reasoning is unknown. The present note details how a straightforward application of Gigerenzer, Hoffrage, & Kleinbölting's (1991) Probabilistic Mental Model (PMM) to the Linda problem provides the first such descriptive account. In PMM terms, the objects are stereotype schemas for each of the societal roles, such as bank-teller and feminist. The cues are the various descriptive personal characteristics, such as outspoken, philosophy major, interested in social justice. The cue validity is a dichotomous variable indicating whether or not a cue applies to Linda. The target variable is an indicator of whether Linda fulfills a particular role or not. Stereotype schemas consist of universal affirmative statements such as "All bank-tellers are college graduates" and "All feminists are concerned with social justice." Stereotype schemas are combined according to standard deductive logic to form conjunctions. As a result, conjunctive schemas are chosen over simple schemas.
Availability: Due to the pending revision, hardcopy reprints are available only on a very limited basis.
This paper was written a few years ago and is awaiting revision. It gives a normative theory of abductive reasoning, relating Peirce's theory to the notion of semantic information by way of a formal proof. The paper is targeted for those interested in epistemology and semantic information.
Abstract: The notion of inference to the best explanation is due to Peirce, who offered a specific formal theory of abductive inference. An important concern regarding Peirce's model is that Peirce provides only an informal model of "bestness," what it means to be the best of a set of deductively equivalent explanations. The present paper offers a positive theoretical account that proposes a resolution to this issue and discusses implications of the resultant new theory of abductive inference. The criterion of bestness suggested is that the best explanation is that which makes for the most natural deductive account of the evidence. Johnson-Laird & Byrne's first criterion for psychological acceptability of deductive arguments is used to determine the most natural account. The deductive argument that throws away the least semantic information is assumed to make for the best explanation. Using this criterion, an expected value for making good abductive inferences is calculated, an algorithm for calculating the best explanation is provided, and an example of statistical estimation of the best explanation under conditions of imperfect information is given. Finally, some of the traditional concerns voiced about explanatory, non-deductive, non-probabilistic inference are discussed in the light of this new model.
Availability: Due to the pending revision, hardcopy reprints are available only on a very limited basis.
This paper is being researched. A target journal has not yet been selected. A series of experiments is being conducted to demonstrate the normative inferential value of the tendency in human reasoning known as the conjunction "fallacy." The effort is to demonstrate that, under realistic conditions, with real base rates for real events, this tendency generates true answers more often than predicted by chance alone. The experiments are proceeding under the erstwhile direction of Anna Romero.
University of North Carolina at Chapel Hill
and Anna Romero
Abstract: The Conjunction Fallacy is the most compelling of the judgmental biases discovered by Tversky & Kahneman. The common judgement made by subjects is that the conjunction is more probable than one of the component propositions. The judgement is compelling, even to the statistically sophisticated and yet it appears to violate any formal notion of probabilistic inference. Both Cohen and Gigerenzer have argued that the various biases and heuristics found by Tversky & Kahneman are not logical fallacies at all. Even on the most generous reading of their arguments against the rest of the biases, the case against the Conjunction Fallacy appears not to have been made. Using Gigerenzer's Probabilistic Mental Model, an argument that the Conjunction Effect is not a fallacy is constructed. If the conjunction effect is the result of a valid form of inference, subjects reasoning so as to produce the conjunction effect should reach true conclusions more often than by random guessing. A series of experiments was performed where subjects were presented with questions about real US cities. In some cases the conjunction gave a true statement about the city, not in others. Subjects were predicted to conclude with conjunctions more often when the conjunction was true than when it was false.
Availability: This paper is in preparation. When drafts are available, the announcement will be made here.
This paper was submitted to the annual meeting of the Association for Behavior Analysis in Atlanta in 1991. It was accepted but never presented due to illness. It is an attempt to relate Peirce's three forms of logical inference to three potentially reinforcing consequences. It is directed towards a behavior analytic audience but should be broadly accessible by anyone interested in a possible relation between inference and reinforcement. When time permits, a revised version is planned for publication.
Abstract: How does one begin to address the question of reasoning within a behavioral analytic framework? Reasoning is a complex behavior that resembles other verbal behavior in certain respects and also resembles the sorts of complex behaviors called "cognitive" by some researchers. As a behavior, reasoning presents two puzzles: 1) What keeps it going? 2) Why are its effects so permanent? I believe that the answer to both of these questions is reinforcement. The key to a Behavioral Analysis of Reasoning is that the various types of reasoning: deductive, inductive, et cetera, are not to be distinguished by virtue of the discriminative stimuli that control them, as one might expect, nor of the response topographies involved, but by virtue of the different contingent reinforcers. In particular, as Catania (1990) suggests, the reward of reasoning is an increased ability on the part of the reasoner to do new things. Reasoning is different from the so-called cognitive behaviors in one crucial respect. After one has reasoned, one's beliefs (that is, one's verbal repetoire) are changed in a more or less enduring fashion. Long term change is the mark of learning. I will argue that this is exactly what reasoning is: learning. And, somewhat surprisingly, in view of the claims of the researchers in respondent conditioning vis a vis cognition, operant learning. Ryle (1949) distinguishes knowing-how from knowing-that. I will argue that, parallel to this distinction, there exists learning-how and learning-that. Learning-how, selection-by-consequences, produces contingency-shaped behavior. Learning-that, reasoning, produces rule-governed behavior. Following Peirce, I argue that there are three types of reasoning: deductive, inductive, and abductive (explanatory). Each of these three generates a type of rule governed behavior well-known to behavior analysts. Deductive reasoning allows predictions. Inductive reasoning generates rules. Abductive reasoning allows rules to govern later behavior. The three types of reasoning are distinguished in virtue of being maintained by three different reinforcers. Evidence in the behavior analytic literature is adduced to support the above claims. Novel experiments are proposed to provide an experimental analysis of reasoning. Finally, implications for applied behavior analysis and behavioral therapies are suggested.
Availability: Due to the pending revision, hardcopy reprints are available only on a very limited basis.
This paper was written a few years ago and is awaiting revision. It attempts to assess Chomsky's arguments against Skinner's approach to language. It is targeted for readers of all backgrounds interested in such topics.
Abstract: A key point in Chomsky's (1959) well-known critique of Skinner's (1957) book, Verbal Behavior, is Chomsky's argument that Skinner's theoretical vocabulary of stimulus, response, and reinforcement, etc. is not extensible from the animal learning laboratory to the everyday life of homo sapiens. It is argued here that Chomsky mis-specified the implications of Skinner's definitions and that an appropriate specification in Skinner's own terms allows for the possibility of extending Skinner's radical behaviorist vocabulary. The impact of this correction for the remainder of Chomsky's arguments is assessed.
Availability: Due to the pending revision, hardcopy reprints are available only on a very limited basis.
This paper is a reprint of an article co-authored with Tom Wallsten and others and published in the Bulletin of the Psychonomic Society.
University of North Carolina at Chapel Hill
David V. Budescu
The University of Haifa
The Pennsylvania State University
and Steven M. Kemp
Abstract: Despite much disagreement regarding how probabilistic information is best communicated, virtually no research has sought to determine what communication modes people prefer or the factors that affect their communication preferences. We report a survey on these issues of 442 graduate and undergraduate students in several specialties and universities. Some group differences emerged, but overall 34% expressed preference for both conveying and receiving information about uncertainty in numerical rather than verbal form, 30% expressed the opposite preferences, and 35% indicated they prefer to receive such information numerically but to convey it verbally. Generally, respondents endorsing the use of verbal information said that it is easier to use, more natural and personal. Those preferring numerical information said that it is more precise. Virtually all respondents, however, evidenced a willingness to use the opposite of their initially preferred mode if the situation warrants it. The willingness to switch from one mode to another was said to depend on the level of precision implied by the data and the importance of the issue, as was suggested by Budescu & Wallsten (1987). These results may be helpful in structuring risk communication strategies.
Availability: This journal is relatively easy to get ahold of. The first author should be contacted for reprints. The full reference is:
Wallsten, T. S., Budescu, D. V., Zwick, R., & Kemp, S. M. (1993). Preferences and reasons for communicating probabilistic information in verbal or numerical terms. Bulletin of the Psychonomic Society, 31, 135-138.
This paper is published as Research Report number 92-1 from the L. L. Thurstone Psychometric Laboratory and can be obtained by contacting them directly. The report was also presented as a poster at the annual meeting of the American Psychological Association in Washington, DC in 1992. The effort was to demonstrate the application of recently available IRT techniques to the analysis of Minnesota Multiphasic Personality Inventory (MMPI) personality scales. As can be seen from the abstract, the underlying project was a very large one. The four authors (including myself) designed an IRT analysis of all of the items with respect to all of the basic MMPI-2 scales. Those data are available for researchers interested in performing further analyses from any of the authors.
Abstract: The item responses of 2,600 "normal" subjects to the revised Minnesota Multiphasic Personality Inventory (MMPI-2) Depression scale were analyzed using the two-parameter logistic Item Response Theory (IRT) model. Parameters were estimated for all items simultaneously. Parameters were also estimated separately for the subsets of items in the Harris-Lingoes and Wiener-Harmon subscales. The dimensionalities of the scale and subscales are investigated, and the strongest dimensions characterized. Contrasts between patterns of responses for men and women are highlighted.
Availability: Contact the L. L. Thurstone Psychometric Laboratory for copies of the research report. Presuming that the appropriate permissions can be obtained, a link for downloading the database will be added here as part of a future update to this WebSite.
This paper was also presented as a talk at the annual meeting of the American Psychological Association in Totonto, OT in 1993. The effort was to demonstrate an application of recently available IRT techniques to questions about the validity of the basic MMPI-2 scales. A summary was published in the January, 1994, issue of The Score, the newsletter for Division 5 of the American Psychological Association. A manuscript version was submitted to a journal and was rejected. The manuscript is awaiting revision.
Abstract: Subtle items are MMPI items whose content is not suggestive of any pathology, despite the fact that these items are each scored on at least one clinical scale. Some (Jackson, 1971) contend that MMPI items scored on clinical scales whose content has no obvious connection to that clinical scale were included with that scale due to psychometric error in the initial construction of the MMPI. For instance, these researchers question the validity of any item scored on the Depression (D) scale, but indicative of no obvious depressive pathology. However, recent investigations (Hovanitz & Jordan-Brown, 1986) of both internal consistency and discriminative validity suggest that subtle items may have use in the MMPI despite their lack of face validity.
Item Response Theory (IRT) provides a new opportunity to examine issues surrounding the subtle/obvious distinction. The slope parameter (a) of an Item Characteristic Curve (ICC) indicates how well a particular item discriminates amongst individuals possessing different levels of the construct (theta) measured by all the items in the scale. Given the assumptions of the IRT model, the slope can be thought of as a measure of the relative contribution of an individual item to the scales ability to measure the construct of interest.
Contributions of subtle and obvious items to the Basic Clinical Scales of the Minnesota Multiphasic Personality Inventory (MMPI) are examined using Item Response Theory (IRT) analysis. Using IRT, JacksonŐs (1971) claim as to the invalidity of subtle MMPI items can be restated as subtle MMPI items should be poor discriminators. For each scale, item discrimination parameters were regressed on Christian, Burkhart, & GyntherŐs (1978) obviousness ratings. Results supported JacksonŐs claim. Correlations showed that for all the basic scales intended to measure major psychological disorder, obviousness is a good predictor of discrimination. Correlations by sex on scale 3 (Conversion Hysteria) suggest different factor structures for males and females. Item content was compared to ComreyŐs (1957) factor interpretations of the Hy scale. Across the basic scales, items lacking obvious content apparently fail to load on the same factor with the obvious items.
Availability: The pending revision of this paper is unlikely to happen any time soon. Therefore, hardcopy reprints are available on a limited basis.
This paper was an invited talk for the annual meeting of the Association for Behavior Analysis in Atlanta in 1991. It was never presented due to illness. It is an compilation of quotations by Skinner and Peirce, attempting to relate their views. It is directed towards a behavior analytic audience but should be broadly accessible by anyone interested in the philosophy of mind. When time permits, a revised version is planned for publication.
Abstract: A number of researchers have begun to consider the conceptual similarities between Radical Behaviorism and the philosophy of Pragmatism (see Lee, 1988 for examples), in particular the pragmatism of John Dewey. This paper will consider intriguing parallels between the philosophy of B. F. Skinner, founder of the school of Radical Behaviorism, and C. S. Peirce, founder of the school of Pragmatism. The relevant positions taken by each can be collected (loosely) into three groups:
As regards their overall approach to science: both were committed Darwinians; both sought to explain and classify phenomena in terms of future effects, rather than past causes; both were skeptical of the importance of reference to meaning, preferring a more pragmatic conception; and both conceived of explanation in terms of future progress in scientific inquiry, rather than in the demonstration of the existence of putative causes.
Further, many of the views generally taken to exemplify Skinner's position as distinct from other philosophies of psychology were views shared by Peirce. Both took a "black box" approach to the problems of mind, but neither denied the existence of internal goings-on, merely that the way to find out about such matters was to examine only the relations amongst the external phenomena. Both felt that the underlying laws of psychology were to be found in the variations between individual organisms, not in the common variations across groups.
Finally, they shared crucial points of view about metaphysics. Skinner's arguments against Hull's view of reinforcement echo Peirce's arguments against Leibniz' view of predicates. As Flanagan (1980) points out, Skinner's materialism was minimalist with respect to its purpose in his overall philosophy. Peirce's Scotistic realism was minimally distinct from nominalism and served a similar purpose. Skinner considered the first response in an operant class to be a special case. In an analogous fashion, Peirce required abductive inference to bootstrap any inductive reasoning. Finally, both were committed holists.
The relevant positions taken by each are illustrated with quotations. Possible benefits of integrating some of Peirce's approach into contemporary behavior analysis are suggested.
Availability: Due to the pending revision, hardcopy reprints are available only on a very limited basis.
This paper is a preprint of an article submitted to The Behavior Analyst based on a talk I gave at the 1996 annual meeting of the Association for Behavior Analysis. (I was the only theist in a room full of atheists.) It has been returned for revision and resubmission. More accessible to a general analysis than most and, I hope, more fun.
Abstract: Since the ascendancy of science, the compatibility of science and religion has been debated. Both scientific and religious behaviors can be subjected to behavior analysis, but behaviorism itself is based upon scientific principles. Can a behavior analysis of this issue lend any insight into this debate? Two questions were raised to focus the discussion. One, how can a scientist believe in something she or he cannot prove? Two, what effect does a person's belief in God have on his or her behavior? The first question leads to a consideration of both science and religion as cultural systems wherein verbal behavior is used to alter the behavioral propensities of the participants. Differences in the methods used by each system (belief and faith) suggest that there exist participatory stances with respect to each that are compatible with participation in the other, though not simultaneous participation in both. The second question seems to derive from concerns about negative effects of science and religion on society. Technological advances in warfare and religious persecutions as well as numerous other effects allow proponents of one system to derogate the effects of the other. An argument is presented to the effect that neither religion nor science is inherently negative in its impact on behavior and society.
Availability: Currently, the unrevised version is available in hardcopy only. Both html and PostScript versions are planned as soon as the status of the resubmission is determined and the appropriate permissions are obtained.