![Medin, D.L. & Schaffer, M.M. (1978). Context theory of
classification learning. Psychological Review, 85,
207-238.
Merz, C. J., & Murphy, P. M. (1996). UCI Repository of
machine learning databases [http://www.ics.uci.edu/
~mlearn/MLRepository.html]. Irvine, CA: University of
California, Department of Information and Computer
Science.
Miller, C. S. & Laird, J. E. (1996). Accounting for graded
performance within a discrete search framework.
Cognitive Science, 20, 499-537.
Nosofsky, R. M. (1986). Attention, similarity, and the
identification-categorization relationship. Journal of
Experimental Psychology: General, 115(1), 39-57.
Quinlan, J. R. (1993). C4.5: Programs for machine
learning. Los Altos: Morgan Kaufmann.
Restle, F. (1961). Psychology of judgment and choice. New
York: Wiley.
Schlimmer, J.S. (1987). Concept acquistion through
representational adjustment (Tech. Rep. 87-19).
University of California, Irvine. Doctoral dissertation,
Department of Information and Computer Science.
Tversky, A. (1972). Elimination by aspects: A theory of
choice. Psychological Review, 79(4), 281-299.](https://image.slidesharecdn.com/97cogscidoc1958/85/97cogsci-doc-7-320.jpg)
97cogsci.doc
- 1. Berretty, P.M., Todd, P.M., and Blythe, P.W. (1997). Categorization by Elimination: A fast and frugal approach to categorization. In M.G. Shafto and P. Langley (Eds.), Proceedings of the Nineteenth Annual Conference of the Cognitive Science Society (pp. 43-48). Mahwah, NJ: Lawrence Erlbaum Associates. Categorization by Elimination: A Fast and Frugal Approach to Categorization Patricia M. Berretty (berretty@mpipf-muenchen.mpg.de) Peter M. Todd (ptodd@mpipf-muenchen.mpg.de) Philip W. Blythe (blythe@mpipf-muenchen.mpg.de) Center for Adaptive Behavior and Cognition Max Planck Institute for Psychological Research Leopoldstrasse 24, 80802 Munich Germany the rabbit could also use several cues to make its category assignment, as soon as it finds enough cues to decide Abstract “predator”--for instance, that this bird is gliding--it will not bother gathering any more information, and instead will People and other animals are very adept at categorizing head for shelter. Obviously, in the case of the rabbit, speed stimuli even when many features cannot be perceived. is of the essence when categorizing birds as predators or Many psychological models of categorization, on the other nonpredators. Humans face similar circumstances where hand, assume that an entire set of features is known. We present a new model of categorization, called Categorization rapid categorization is called for, making use of only by Elimination, that uses as few features as possible to whatever information is immediately available. Being able make an accurate category assignment. This algorithm to categorize rapidly the intention of another approaching demonstrates that it is possible to have a categorization person as either hostile or courting, for instance, will process that is fast and frugal--using fewer features than enable the proper reactions to ensure the most desirable other categorization methods--yet still highly accurate in its outcome. judgments. We show that Categorization by Elimination In this paper, we consider the case for a “fast and frugal” does as well as human subjects on a multi-feature (à la Gigerenzer & Goldstein, 1996) model of categorization task, judging intention from animate motion, categorization, akin to the lexicographic process of bird and that it does as well as other categorization algorithms identification described in the first paragraph. This model, on data sets from machine learning. Specific predictions of the Categorization by Elimination algorithm, such as the which we call Categorization by Elimination (CBE), uses order of cue use during categorization and the time-course only as many of the available cues or features as are of these decisions, still need to be tested against human necessary to first make a specific categorization. As a performance. consequence, it often uses far fewer cues in categorizing a given stimulus than do the standard cue-combination models, yielding its fast frugality. This information- 1. Introduction processing advantage can be crucial in a variety of categorization contexts where speed is called for, as in Hiking through the Bavarian Alps, you come upon a identifying threats. On the other hand, the accuracy of this large bird gliding over a meadow. You pull out your approach typically rivals that of more computationally European bird guidebook to identify it. From the shape of extensive algorithms, as we will show. We therefore its body, you assume that this is a bird of prey, so you turn propose Categorization by Elimination as a parsimonious to the section on raptors in the guide. To determine the psychological model, as well as a potentially useful exact species, you next use size to narrow down your candidate for applied machine-learning categorization search to a few kinds of hawks; then you use color to tasks. eliminate a couple more species; and finally with one last Categorization by Elimination is closely related to cue--tail length--you can make a unique classification. Tversky’s Elimination by Aspects (EBA) model of choice Using only four cues (or features), you correctly identify (Tversky, 1972). After describing competing this bird as a sparrow hawk. You could take out your psychological and machine-learning models of binoculars and check more cues to support this categorization in the next section, we discuss the identification, but for a rapid decision these few cues are background of elimination models in section 3. We enough. present the Categorization by Elimination model in section How would this categorization process proceed if a 4. Most other recent models of human categorization rabbit rather than a human were watching the bird? The focus on the use of two or three cues, situations in which rabbit would not be interested in knowing the exact species CBE can show little advantage. Therefore, we have of bird flying overhead, but rather would want to experimentally investigated a multiple-cue categorization categorize it as predator or not, as quickly as possible--the task in which we can compare our model with others in Rabbit’s Guide to Birds has only two short sections. While accounting for human performance with seven cue
- 2. dimensions. We describe this study, which involves usually use all of the cues that are present--that is, do not categorizing animate motion trajectories into different discard any available information. behavioral intentions, in section 5. CBE does as well as Exemplar models (Brooks, 1978; Estes, 1986; Medin & linear categorization methods, and does not overfit the data Schaffer, 1978; Nosofsky, 1986) assume that when as neural networks seem to. Next, in section 6 we look at presented with a novel object, humans compute the how well our algorithm does alongside some of the similarity between that object and all the possible multiple-attribute categorization methods developed in categories in which the novel object could be placed. psychology and machine learning on standard data sets Similarity is a function of the sum of the distances between from the latter field. This comparison shows that the object and all the exemplars in the particular category. Categorization by Elimination can often compete in The object is placed into the category with which it is most accuracy with more complex methods. Further, if similar. minimizing the number of cues used is sought to maximize Nosofsky’s (1986) generalized context model (GCM) computational speed, CBE usually emerges as the clear allows for variation in the amount of attention given to winner. Finally in section 7 we consider some of the different features during categorization (see also Medin & challenges still ahead, including how to test CBE against Schaffer, 1978). Therefore, it is possible that different human learning data. cues will be used in different tasks. However, this attention weight remains the same for the entire stimulus 2. Existing Categorization Models set for each particular categorization task, rather than varying across different objects belonging to the same Many different models of categorization have been category (in contrast to our new method, as we will see). proposed in both the psychological and machine learning Decision Bound Theory (or DBT--see Ashby & Gott, literature. Psychologists are primarily concerned with 1988) assumes that there is a multidimensional region developing a model that best describes human associated with each category, and therefore that categories categorization performance, while in machine learning the are separated by bounds. An object is categorized goal is to develop an optimally-performing model--that is, according to the region of perceptual space in which it lies. one with the highest accuracy of categorization. These two Similarly, neural network models (e.g., Kruschke, 1992) goals are not necessarily mutually exclusive; indeed, one learn hyperplane boundaries between categories, capturing of the main findings so far in the field of human this knowledge in their trainable weights. In both cases, all categorization is that people are often able to achieve near of the cues available in a particular stimulus are used to optimal performance (that is, categorize a stimulus set with determine the region of multidimensional space, and hence minimal errors--see Ashby & Maddox, 1992). As a the associated category, in which that stimulus falls. consequence, some models, including neural networks and These psychological models all categorize by integrating SCA (Miller & Laird, 1996) are often aimed at filling both cues and using all the cues available (except in GCM if a roles. cue has an attention weight of 0). In addition, training However, the majority of psychological studies of these models to learn new categories is a relatively simple categorization have used simple stimuli that vary on only a process. But the memory requirements assumed by these few (2-4) dimensions, unlike the typical high-dimensional models do differ: for example, GCM assumes that all machine learning applications. It remains to be seen exemplars ever encountered are stored and used when whether humans can also be optimal at categorizing multi- categorizing a novel object, while DBT does not need to dimensional objects. In addition, the predominant store any exemplars. In comparison, our CBE algorithm psychological models of categorization have not addressed does not integrate all available cues, is similarly easy to the issue of constraints, such as limited time and train, and typically requires little memory. information. What might the categorization process be Another approach to psychological modeling is captured when there are both time and information constraints, in the discrete symbol-processing framework of Miller and either because there is an overwhelming number of Laird’s (1996) Symbolic Concept Acquisition (SCA) possible cues to use or only a subset of cues available? model. Here rules are built up incrementally for Here we briefly review some of the currently popular classifying stimuli according to specific features, categorization models for human categorization and beginning with very general rules that test a single feature machine learning with these questions in mind. and progressing to more detailed rules that must match the Throughout the remainder of the paper we use the terms stimuli on many features. While there are similarities cues, aspects, dimensions, and features, as appropriate, to between this approach and CBE (in particular, the order in all mean roughly the same thing. which features are processed can be related to our cue The predominant theories of categorization in the validity measure), one major difference is that new stimuli psychology literature include exemplar models (Nosofsky, are first checked against rules using all available cues, and 1986), decision bound models (Ashby & Gott, 1988), and only when this fails are fewer cues tested against the more neural network models (e.g. ALCOVE--see Kruschke, general rules. In contrast, CBE begins with a single cue, 1992). Each of these categorization models assumes that and only adds new ones if necessary, thereby minimizing the stimuli may be represented as points in a computation. The earlier EPAM symbolic discrimination- multidimensional space. Furthermore, these models all net model (Feigenbaum & Simon, 1984) tests rules in the assume that humans integrate features--that is, combine multiple cues to come to a final judgment--and that we
- 3. efficient general-to-specific method we advocate, but the important aspects). Possible remaining choices that do not rest of our approach is distinct. possess the current aspect being used for evaluation (for In machine learning, predominant categorization theories instance, restaurants that do not serve seafood) are include neural networks, Classification and Regression eliminated from the choice set. Furthermore, only aspects Trees (or CART--see Breiman, Friedman, Olshen, & that are present in the most recent choice set are considered Stone, 1984), and decision trees (e.g., ID3--see Quinlan, (for instance, if all nearby seafood restaurants are cheap, 1993). The goal of these machine learning models is then expense will not be used as an aspect to distinguish usually to maximize categorization accuracy on a given further among them). Additional aspects are used only useful data set. Algorithm complexity and speed are not until a single choice can be made, which is different from typically the most important factors in developing machine the categorization models described above that use all learning models, so that many are not psychologically cues. plausible. One model that does attempt psychological plausibility by applying selective attention to unsupervised concept 4. Categorization by Elimination formation is Gennari’s (1991) CLASSIT. This system Our new Categorization by Elimination algorithm is a classifies objects initially using a subset of the available noncompensatory lexicographic model of categorization, in cues determined by their attentional salience. However, all that it uses cues in a particular order, and categorization cues must still be considered before a final decision is decisions made by earlier cues cannot be altered (or reached, due to a “worst case” stopping rule. compensated for) by later cues. In CBE, cues are ordered Thus, even though many of the machine learning models and used according to their validity. For our present (e.g., CART and CLASSIT) use only a few cues during a purposes we define validity as a measure of how accurately given categorization, the process of setting up the a single cue categorizes some set of stimuli (i.e., percent algorithm’s decision mechanisms beforehand, including correct). This is calculated by running CBE only using the determining which cues to use, can be very complex. In single cue in question, and seeing how many correct contrast, our CBE algorithm has a simple learning phase, categorizations the algorithm is able to make. (If using the and still maintains comparable accuracy using few cues. single cue results in CBE being unable to decide between multiple categories for a particular stimulus, as will often 3. Elimination Models be the case, the algorithm chooses one of those categories Motivated by the concerns raised in section 1, we at random--in this case, cue validity will be related to a wanted to develop a fast and frugal categorization method cue’s discriminatory power.) Thus if size alone is more that combines the best aspects of both the psychological accurate in categorizing birds (or more successful at and machine learning models. From the psychological narrowing down the possible categories) than shape alone, models we used the concepts of simple training and size would have a higher cue validity than shape. (There decision processes and a small memory load. From the are other ways that cues can be ordered besides by validity, machine learning models we took the notion of such as randomly or in order of salience, which we are categorizing stimuli without using all available cues. This currently exploring.) combination led us to look into elimination models. CBE assumes that cue values are divided up into bins Classical elimination models were conceived of for (either nominal or continuous) which correspond to certain choice tasks (Restle, 1961; Tversky, 1972) In a sequential categories. These bins form the knowledge base that CBE elimination choice model, an object is chosen by uses to map cue values onto the possible corresponding repeatedly eliminating subsets of objects from further categories. As an example, the size cue dimension for consideration, thereby whittling down the set of remaining birds could be divided into three bins: a large size bin possibilities. First a particular subset of the original set is (which could be specified numerically, e.g. “over 50 cm”) chosen with some probability, using a particular feature to corresponding to the categories of eagles, geese, and determine the subset members. Subsequent subsets are swans; a medium size bin corresponding to crows, jays, and chosen in the same manner, with successive features, until hawks; and a small size bin corresponding to sparrows and only one object remains. finches. The most widely known elimination model in To build up the appropriate bin structures, the relevant psychology is Tversky’s (1972) Elimination by Aspects cue dimensions to use must be determined ahead of time. (EBA) model of probabilistic choice. One of the At present we construct a complete bin structure before motivating factors in developing EBA as a normative testing CBE’s categorization performance, but learning and model of choice was that there are often many relevant testing could also be done incrementally. In either case, cues that may be used in choosing among complex bins can be constructed in a variety of ways from the alternatives (Tversky, 1972). Therefore, part of any training examples--in the next two sections, we present two reasonable psychological model of choice should be a possibilities. procedure to select and order the cues to use from among A flowchart of CBE is shown in Figure 1. Given a many alternatives. In EBA, the cues, or aspects, to use are particular stimulus to categorize, an initial set of possible selected according to their utility for some decision (for categories is assumed, along with the ordered list of cue instance, to choose a restaurant from those nearby, what dimensions to be used. The categorization process begins they serve and how much they charge might be the most by using the cue dimension C with the highest validity.
- 4. Next a subset S of the possible categories is created evidence for CBE to come from situations in which containing just those categories that correspond to the first categorization may be affected by time and cue availability cue C’s value for the current stimulus object (this subset is constraints. As mentioned in the introduction, one specific determined through the binning map described earlier). If domain where time and number of available cues are only one category corresponds to that cue value, the limited is in inferring intention from motion. Blythe, categorization process ends with this single category. If Miller, and Todd (1996) conducted an experiment in which more than one category corresponds to the current cue the subjects’ task was to infer intention from motion of two value, that set of possible categories S is kept, and the cue animated bugs shown moving about on a computer screen. with the next highest validity, C*, is checked. The set of The movement patterns had all been previously generated categories S corresponding to the previous cue C’s value is by other subjects instructed to engage in various types of intersected with the set, S*, of categories corresponding to interaction by each controlling the motion of a single on- the present cue C*’s value. This is CBE’s elimination screen bug. The six possible categories of interactive step. motion were: pursuit, evasion, fighting, courting, being If only one category remains in the new set intersection, courted, and playing. For example, one subject’s task the algorithm terminates at this point with that one would be to have their bug pursue the other bug, while the category. If more than one category remains, this other subject would move their bug to evade their pursuing intersection becomes the new set S of remaining opponent. Next, new subjects viewed the recorded bug possibilities, the next cue is checked, and the process is interactions and through forced choice, categorized the repeated. If the intersection is empty, then the present cue interactions as a specific type of intentional motion. is ignored, the prior set S of categories is retained, and the Seven salient cues of motion were calculated for each of next cue is evaluated. This process of checking cues and the recorded motion patterns (see Blythe, Miller, & Todd, using them to reduce the remaining set of possible 1996, for details). These cues were used to compare categories continues until a single category remains, or different categorization models with each other and against until all the cues have been checked, in which case a human performance. (While we cannot be sure that these category is chosen from the remaining set at random. are the exact cues used by the human subjects, it is a This algorithm has several interesting features. It is reasonable set to start with.) frugal in information, using only those cues necessary to The four models tested were CBE and three traditional reach a decision. It is non-compensatory, with earlier cues cue-integrating compensatory algorithms: unit tallying eliminating category-choice possibilities that can never be (counting up the total number of cues that indicate one replaced by later cues. The binning functions used to category versus another, using the same bin mapping as associate possible categories with particular cue values can CBE), weighted tallying (adding up weighted votes from be as simple or detailed as desired, from one-parameter median cuts to multiple-cutoff mappings. And the exact Get first cue C order of cues used does not appear to be critical: in preliminary tests, different random cue orderings vary the algorithm’s categorization accuracy by only a few Set possible cat.s percentage points (but, interestingly, the number of cues S to categories used with different orderings does vary more widely). corresponding to CBE is clearly similar to EBA in several aspects, though prev. cue C’s value there are some important differences. First, EBA is a probabilistic model of choice while CBE is (in its current form) a deterministic model of categorization. Second, in Get next cue C* CBE cues are ordered before categorizing so that the same cue order is used to evaluate each object. In EBA, aspects are selected probabilistically according to their weight. Set possible cat.s Therefore, the order of aspects is not necessarily the same S* to categories for each object. Third, as mentioned previously, EBA only corresponding to chooses aspects that are present in the current set of current cue C*’s value remaining possible choices, and therefore the process never terminates with the empty set. However, to select such an aspect, all candidates must be examined to determine Set S to which aspects are still possible to use. CBE does no such if S contains intersection of if S is checking ahead of time for appropriate cues to use, but more than 1 cat. S and S* empty rather takes this circumstance into account in its behavior when the intersection of current and previous possible if S contains category sets comes up empty. 1 category STOP 5. CBE and Human Data Under what conditions might CBE be a plausible Figure 1: Flow Diagram of CBE description of human categorization? We expect the most
- 5. all the cues that indicate one category versus another, again and information-availability constraints are taken into using CBE’s binning along with weights determined by account. correlation), and a three-layer feed-forward neural network model trained by backpropagation learning (see Gigerenzer 6. CBE and Machine Learning Algorithms & Goldstein, 1996, for more details on the first two). It is difficult to compare CBE to existing categorization The bin structure used for CBE and the tallying models on multiple-cue human data beyond the domain algorithms were determined by considering the distribution just presented, because few other experiments have been of cue values for each category and placing the bin performed with more than three or four cues. Instead, as boundaries at points of minimum overlap between an alternative test of CBE’s general accuracy potential, we categories. As a result, some cue values could be mapped examined how well CBE categorized various multi- onto too few possible categories (e.g. if pursuit was usually dimensional objects using data from the UCI Machine fast and courtship usually slow, the fast velocity bin would Learning Repository (Merz & Murphy, 1996). We only map to pursuit, and thus would miss all those compared the performance of CBE to a standard exemplar instances model and a three-layer feed-forward neural network Method Cat Acc Avg trained with backpropagation. Results are shown in Table Cues 2 for categorization performance when trained on the full Subject 49.33% ? data sets and generalization performance when trained on CBE 65.33% 3.77 half of each data set and tested on the other half. In Wtally 64% 7 addition, we include the best reported categorization Utally 63.33% 7 performance we have found for each data set in the Nnet 88.33% 7 machine learning literature. For the following comparisons, CBE used “perfect” Table 1: Categorization accuracies and average number of binnings for the cue values. This means that the cue-value cues used for subjects and models (chance = 16.7%). bins always map to the entire set of possible categories associated with those particular cue values (unlike the bins of rapid, excited courtship motion). Thus this bin mapping in the motion categorization example, where only the most made perfect categorization impossible for CBE in this prevalent categories in each bin were returned). With domain, and yet it did surprisingly well. Table 1 lists the perfect binning, the same categorization accuracy is always average categorization accuracies of the human subjects, obtained regardless of the order in which the cues are used. the categorization accuracies for the four models, and the However, when the cues are ordered by validity, average number of cues used (this value is unknown for categorizations can be accomplished using the fewest the human subjects). Since there were six possible number of cues. categories, chance performance is 16.7%. Table 2 shows the results of these comparisons for three As can be seen in Table 1, human subjects performed data sets. The first is the famous iris flower database in well above chance in this task, and the four categorization which there are 150 instances classified into three algorithms performed better still. The neural network did categories (different iris species) using four continuous- suspiciously far better than the human subjects, indicating valued features (lengths and widths of flower parts). The that it has possibly been overtrained on this data. (When next comparison used wine recognition data, in which 13 tested on generalization ability on a further untrained set of chemical-content cues are used to classify 178 wines as motion stimuli, the network’s performance drops to 68%, one of three particular Italian vintages. The third data set while the other three algorithms hover around 56%.) The analyzed contains two mushroom categories, poisonous tallying algorithms and CBE are all much closer to human and edible, with 22 nominally valued dimensions, and performance, but CBE achieves its accuracy while using 8124 total instances. only about half of the cues of the others. Overall, CBE does very well on these three sets of multi- The difference in accuracy between subjects and the feature natural objects, while using only a small proportion algorithms can be explained in part by the fact that the of the available cues. CBE even performs well in algorithms are “trained” on all the stimuli, either through comparison with models specifically designed for the best the binning process or neural network learning (300 motion possible performance on these particular data sets. We patterns in this case). In contrast, subjects must make their were not expecting CBE to outperform these specialized categorizations without previous exposure to these stimuli algorithms; merely being in the same ballpark is a (under the assumption that they would already know the powerful testament to this approach’s potential accuracy cue structures of these categories through their experiences across varied domains. Furthermore, these algorithms all outside the lab). To make a more fine-grained assessment employ the full set of cues, making the contrast with of how well each categorization algorithm matches the CBE’s high accuracy through limited information all the human data, we are performing analyses of the case-by- more striking. case categorizations made by subjects and algorithms. But even without this detailed analysis, CBE emerges as a 7. Future Work parsimonious contender among categorization algorithms in this multi-cue domain, and the clear winner when time The results we have presented here indicate that a fast and frugal approach to categorization is a viable alternative
- 6. to cue-integrating compensatory models. By only using knowledge, rivaling the abilities of much more complex those cues necessary to first make a categorical decision, and sophisticated algorithms (not to mention human CBE can categorize stimuli under time pressure and subjects!). information constraints. Moreover, if certain cues are The following issues still need to be explored. First, missing (i.e. some feature values are unknown or cannot be how should bin structures be created? Incremental perceived), CBE can still use the other available cues to learning can build the cue-value bins gradually as more come up with a category judgment (we are in process of and more stimuli are seen. But how far should this collecting data on this type of generalization ability across learning process go, and in what way should it proceed? different categorization algorithms). Yet CBE still We have presented two alternatives here, and there are performs very accurately, despite its limited use of many others possible. One Table 2: Categorization accuracies and average number of cues used for various models on three data sets Model Train/Test Iris Wine Mushroom Set Size Cat Acc Avg Cues Cat Acc Avg Cues Cat Acc Avg Cues CBE Full 91.33 % 40.00 % 96.63 % 20.74 % 91.71 % 26.11 % Half 92.40 % 26.24 % 90.37 % 15.83 % 91.66 % 26.16 % Nnet Full 97.67 % 100 % 100 % 100 % 86.21 % 100 % Half 97.07 % 100 % 95.95 % 100 % Best 98 % 100 % 95.00 % Reported (James, 1985) (Aeberhard et al., 1992) (Schlimmer, 1987) important issue to explore further is the performance stimuli. Journal of Experimental Psychology: Learning, tradeoff between accuracy and the amount of knowledge Memory, and Cognition, 14, 33-53. captured in the bin structure (CBE’s memory Ashby, F.G. & Maddox, W.T. (1992). Complex decision requirements). rules in categorization: Contrasting novice and exper- Second, more data from human performance on ienced performance. Journal of Experimental categorizing multi-dimensional objects needs to be Psychology: Human Perception and Performance, 18, collected and analyzed to compare CBE with other 50-71. categorization models. We are particularly interested in Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. investigating the patterns of misclassifications, learning (1984). Classification and regression trees. New York, curves, and predicted time-courses associated with CBE NY: Chapman & Hall, Inc. and human performance. The intriguing finding in our Blythe, P. W., Miller, G. F., & Todd, P. M. (1996). Human intention from motion data that categorization accuracy simulation of adaptive behavior: Interactive studies of varied little with changes in cue order can also be studied pursuit, evasion, courtship, fighting, and play. experimentally. Proceedings of the Fourth International Conference on Third, category base-rates and payoffs for right and Simulation of Adaptive Behavior (pp. 13-22). wrong classifications should be incorporated into the Cambridge, MA: MIT Press/Bradford Books. model. For example, with the mushroom categories Brooks, L. R. (1978). Non-analytic concept formation and described in the previous section, if a mushroom remains memory for instances. In E. Rosch & B.B. Lloyd (Eds.), uncategorized as poisonous or safe even after all the cues Cognition and Categorization, (pp. 169-211) Hillsdale, have been used, it seems reasonable to err on the side of NJ: Erlbaum. caution and guess that the mushroom is poisonous. Estes, W. K. (1986). Array models for category learning. With these further explorations and extensions to CBE, Cognitive Psychology, 18, 500-549. we will come to understand the algorithm’s behavior Feigenbaum, E. A., & Simon, H.A. (1984). EPAM-like better, and be able to make it a better model of human models of recognition and learning. Cognitive Science, behavior in turn. For now, though, we have shown 8, 305-336. evidence for the view that the mind need not amass and Gennari, J. H. (1991). Concept formation and attention. In combine all the available cues when telling a hawk from a Proceedings of the Thirteenth Annual Conference of the dove, or a threat from a flirt--fast and frugal does the trick. Cognitive Society (pp. 724-728). Hillsdale, NJ: Erlbaum. References Gigerenzer, G. & Goldstein, D. G., (1996). Reasoning the Aeberhard, S., Coomans, D., & de Vel, O. (1992). fast and frugal way: Models of bounded rationality. Comparison of classifiers in high dimensional settings Psychological Review, 103(4), 650-669. (Tech. Rep. 92-02). James Cook University of North Kruschke, J. K. (1992). ALCOVE: An exemplar-based Queensland, Dept. of Computer Science and Dept. of connectionist model of category learning. Psychological Mathematics and Statistics. Review, 99, 22-44. Ashby, F.G., & Gott, R. (1988). Decision rules in the perception and categorization of multidimensional
- 7. Medin, D.L. & Schaffer, M.M. (1978). Context theory of classification learning. Psychological Review, 85, 207-238. Merz, C. J., & Murphy, P. M. (1996). UCI Repository of machine learning databases [http://www.ics.uci.edu/ ~mlearn/MLRepository.html]. Irvine, CA: University of California, Department of Information and Computer Science. Miller, C. S. & Laird, J. E. (1996). Accounting for graded performance within a discrete search framework. Cognitive Science, 20, 499-537. Nosofsky, R. M. (1986). Attention, similarity, and the identification-categorization relationship. Journal of Experimental Psychology: General, 115(1), 39-57. Quinlan, J. R. (1993). C4.5: Programs for machine learning. Los Altos: Morgan Kaufmann. Restle, F. (1961). Psychology of judgment and choice. New York: Wiley. Schlimmer, J.S. (1987). Concept acquistion through representational adjustment (Tech. Rep. 87-19). University of California, Irvine. Doctoral dissertation, Department of Information and Computer Science. Tversky, A. (1972). Elimination by aspects: A theory of choice. Psychological Review, 79(4), 281-299.