Correspondence may be addressed to: Janet L. Kolodner, College of Computing, Georgia Institute of Technology, 801 Atlantic, Atlanta, GA 30332-0280. Electronic mail may be sent to jlk@cc.gatech.edu.

In the Chicago science museum, visitors learn about sickle-cell anemia by analyzing the blood of and interacting with simulated clients. Nearby, high school students learn about history by writing and performing newscasts of the events. In a sixth-grade class near Atlanta, students learn about the respiratory system by designing artificial lungs. In a seventh-grade classroom, students learn about arthropods by designing useful arthropod-like robots. Other middle school students are learning about nutrition and logistics by planning a two-week hike on the Appalachian Trail. In a midwestern phone company, employees learn to sell Yellow Pages ads by interacting with simulated clients. In a nearby company, new managers learn strategic planning by solving a strategic planning problem with access to a set of company employees who play the roles they would take on in a real company planning exercise. In an interdisciplinary class at the Georgia Institute of Technology, students learn how to be designers by designing kites and kiosks.

All of these scenarios are real examples of students learning by carrying out authentic and realistic problem-solving activities. The student is given a significant and meaningful mission to carry out, and learning is in the context of carrying out that mission. Modern educational theory stemming from research in the cognitive sciences indicates that knowledge gained through activity that is motivating and authentic is learned more deeply and is more usable than is knowledge gained through memorization, prescriptive activities, or word problems (e. g. , Brown, 1988; Cognition and Technology Group at Vanderbilt, 1993). It is particularly important for students to work on projects that are interesting to them, that ask them to identify the knowledge they need to successfully complete the project, and that require them to apply the knowledge they are (supposed to be) learning (Blumenfeld et al. , 1991; Collins & Brown, 1988; Kolodner, 1993; Schank, Fano, Bell, & Jona, 1993). An important prerequisite to deep and effective learning, this research tells us, is knowing why one needs to learn something. Students need to reflect appropriately on their experiences so as to be able to extract and articulate what they have learned. Each of the scenarios above has all of these properties.

In addition, each scenario is guided by principles stemming from the study of analogical and case-based reasoning (CBR; cf. Gentner & Markman, 1997, this issue; Holyoak & Thagard, 1997, this issue; Kolodner, 1993; Riesbeck & Schank, 1989). For example, in each scenario, an important resource for students as they are solving their problems is the experiences of others. Students learning about sickle-cell anemia have available to them video clips of genetic counselors who give them advice in the form of stories, and those putting newscasts together have access to feedback (in the form of videotaped stories about their own experiences) from expert journalists, historians, and newscasters. Students designing artificial lungs have access to an on-line library of artificial organs and how they work. Those planning a hike on the Appalachian Trail have access to the stories of other hikers (on the World Wide Web), and so on.

The intention is to provide useful cases and examples to students as they are solving problems to enable them to make useful analogical inferences: to identify issues to pay attention to, to form ideas about how to move forward, and to project the effects of solutions they have come up with. The analogy literature tells us that reasoners naturally use their own experiences for such reasoning (e. g. , Klein, Whitaker, & King, 1988; Kochen, 1983; Read & Cesa, 1990; Ross, 1986, 1989). As novices, however, students might not have previously had the most relevant experiences. In these classroom situations, their own experiences are augmented, in effect, by those of others to enable them to reason beyond what they could do based only on what they already know. The teacher as a facilitator (and in some cases, the computer playing that same role) helps students make necessary mappings between old cases and the new situation, if students are having trouble doing that.

Students in each of these situations also spend time interpreting the success or failure of their ideas and reflecting on what they have learned and when they might find those lessons relevant in the future. Built into the curriculum is the reflection needed to promote analysis and encoding of students’ experiences in ways that will make them useful and accessible in the future at opportune times. Although it is impossible to encode one's experiences in ways that guarantee that they will be retrieved at the right times later, research on memory retrieval from both the artificial intelligence (Hammond, 1989; Kolodner, 1983; Owens, 1993; Schank, 1982) and psychology (Patalano, Seifert, & Hammond, 1993; Seifert, 1989) literatures suggests that one can enhance one's ability to recall by anticipating the situations in which a lesson learned might be usefully applied. And lessons of practice in building computer systems that reason on the basis of their experiences provide guidelines on the kinds of content that ought to be encoded for best reuse of experience (Kolodner, 1993).

The practices and software used in these situations are based broadly on the research in analogical reasoning discussed in the two companion articles in this issue (Gentner & Markman, 1997; Holyoak & Thagard, 1997) and more specifically on a particular approach to analogical reasoning called CBR (Hammond, 1989; Kolodner, 1993; Kolodner & Simpson, 1989; Riesbeck & Schank, 1989). CBR focuses on reasoning that is based on previous experience. A previous experience might suggest a solution to a new problem or a way of interpreting a situation, may warn of a problem that will arise, or may allow the potential effects of a proposed solution to be predicted—all types of inferences necessary for addressing the kinds of ill-defined or complex problems that come our way in the workplace, at school, and at home.

CBR has as its centerpieces (a) analogy in the context of solving real-world problems and (b) a research methodology of computational modeling aimed at deriving hypotheses about cognition. Its emphasis on analogy in the context of real-world problem solving has led CBR researchers to focus on reasoning that is based on nearly matching analogs, as those are the type of analogs used most extensively, easily, and successfully by people (Faries & Reiser, 1988; Klein et al. , 1988; Ross, 1986, 1989). The focus has been on analogs representing personally experienced situations. Such analogs, called cases, include in them a rich representation of an old situation and its problem, the way the situation was dealt with, and the results of dealing with it in that way.

CBR was originally developed (in the early to mid 1980s) as a methodology for enhancing computer cognition. Its aim was to allow development of more capable expert systems, ones that could become more capable and efficient reasoners through experience. The idea was that (a) the computer solves problems by remembering old situations and basing its decisions on those, and (b) each problem encountered is entered into memory, resulting (sometimes) in learning new ways of solving problems, new ways of describing problems, and new paths for access (indexing).

CBR's computational models, developed to implement such systems, were originally based on both observations of people and on the analogical reasoning literature that has focused on people solving problems in the world (e. g. , Faries & Reiser, 1988; Gilovich, 1981; Klein et al. , 1988; Read & Cesa, 1990; Ross, 1986). But its emphasis is different than that of traditional analogical reasoning research. Because of CBR's focus on the uses of analogs in everyday reasoning and its desire to build reasoning systems that can carry through in solving whole problems, CBR's research focus has been on encoding, retrieval, and use of analogs and on plausible interactions between memory, reasoning, and learning processes. Because of its focus on near analogs, CBR is less directly concerned with mapping. Its focus on computational modeling as a research paradigm means that the CBR community sees its role as providing plausible hypotheses about cognition, leaving the gathering of empirical evidence on its hypotheses to those who are trained at psychological research.

Indeed, the computational models CBR has put forth of encoding, retrieval, and adaptation have inspired the analogical reasoning community to take encoding and retrieval seriously (e. g. , Gentner & Forbus, 1991; Lange & Wharton, 1993; Thagard, Holyoak, Nelson, & Gochfeld, 1990) and have inspired a variety of empirical investigations (Patalano et al. , 1993; Seifert, 1989) of analogical reasoning as it happens in the wild. In addition, its algorithms provide insight into what it might take to enhance human cognition. CBR as a plausible cognitive model can thus advise on educational philosophy, educational practice, and design of software (Kolodner, Narayanan, & Hmelo, 1996; Schank & Cleary, 1995).

In this article, I walk through an example of CBR in practice and point out several of its characteristics, discuss the cognitive and computational models CBR implies and its contributions to our understanding of remembering and learning, and consider its implications for education and for the design of educational software and provide some examples of its ideas in use.

An architect is designing an office building with a long, naturally lit atrium in the center and a circular row of offices surrounding it. ¹ To minimize energy consumption, she wants the office to get as much light as possible. She remembers an existing library in which the designer solved the problem of bringing in sunlight by constructing exterior walls of glass. She considers this for the current situation: The office space can be separated from the atrium by a circular glass wall. But then she remembers the problems that a courthouse had. There, a glass wall was used in a row of offices with heavy public traffic. Although the offices were well lit, the constant presence of the public interfered with the privacy and work of the office workers. The library had not had this problem, because the glass wall faced a wooded area. But this second case alerts her to a potential problem with the solution she has derived. She considers whether it would indeed be a problem for the office building. Comparing the two recalled cases with the current one, she realizes that the potential for this problem exists but to a lesser degree than in the courthouse. Although the atrium is not deserted like the woods, it is not a heavily trafficked area either. She decides to use the glass-wall solution but modifies it slightly by using translucent glass bricks instead of clear plate glass, keeping the small amount of traffic that there is from annoying those in the adjoining offices.

This example shows a reasoner applying knowledge from specific examples to solving a new problem. One case suggested a solution to a problem and another warned about a potential problem; the reasoner was able to solve her problem adeptly based on both kinds of suggestions. The example also illustrates qualities that analogs must have to be useful in real-world problem solving. The architect could not have come to the decision to use glass bricks in the wall had she not known about the failure of the glass wall in the courthouse and had available an explanation of that failure. The better the analysis of successful or failed goal achievement, the more useful a case will be.

As the example shows, CBR focuses on and makes explicit the role of personal experience in learning. Cases record both positive and negative experiences—successes and failures at accomplishing goals. Remembering an experience can suggest a way to do something, what to focus on, or what to avoid. Thus, a case is useful to the extent that it holds useful content and is accessible at the right times.

The above example also shows that a case is more than just an instance or an example; it is the interpreted representation of a real experience. Thus, it includes a sought-after goal, a method for achieving the goal or solution to the problem, and the results (outcome) of carrying out that method (solution), all of this described specifically. A case that records explanations of why an experience succeeded or failed helps a reasoner to discern when strategies used previously should be repeated or avoided.

This example is similar to others in which people use CBR in both commonsense and expert domains. Researchers at GTE found that engineers who were reasoning about what could go wrong with phone switching networks used cases extensively (Kopeikina, Brandau, & Lemmon, 1988). Our own observations of car mechanics have shown that both novice and experienced mechanics use their own experiences and those of others to generate hypotheses about what is wrong with a car and to remember how to test for different diagnoses (Lancaster & Kolodner, 1987). Klein and Calderwood (1988) observed expert decision makers in several different naturalistic, complex, dynamically changing situations using near analogs to understand situational dynamics, generate options, predict the effects of implementing an option, and reassure themselves of the reliability of an option. The primary power of cases in these situations, claimed Klein (Klein et al. , 1988), is that it allows the decision maker to deal with the unknown and uncertain information. Read and Cesa (1990) observed people using old cases for explanation of anomalous occurrences and found them particularly adept at doing that when the anomalous event reminded them of a personal experience.

Ross (1986, 1989) showed that people learning a new skill often refer back to previous problems to refresh their memories on how to do the task. He also identified difficulties novices have with using their experiences, because they lack a good understanding of the domain:

• They often have not encoded their experiences well, making retrieval at appropriate times difficult.

• They have trouble mapping, preferring to map similar components to each other rather than components that play the same role.

• They cannot reuse experiences when they have not had the relevant experiences to recall.

These studies show that reasoning using analogs is a natural process for people, especially when there is much uncertainty or many unknowns and during early learning, and they suggest some of the aid we should give to novice learners if we want to help them reuse their experiences well.

The process of carrying out CBR includes, first, retrieval, in which an analog is found in memory. This may be easy, as one's description of a new situation might match some experience already in memory. Or, it may require reinterpretation and rerepresenting a situation in different or more specific terms or from a different point of view. Such reinterpretation might be done incrementally, creating better and better descriptions of a situation on the basis of what is recalled or not recalled in earlier probes and its usefulness.

Once a case or cases are retrieved, they can be used. An old solution might be adapted to solve a new problem, pieces of several old situations might be merged to create a new solution, predictions might be made based on an old situation, an old and new situation might be compared and contrasted to determine important issues to focus on or what needs to be adapted, and so on. The next step is action that is based on inferences, which leads to results. If results are different than expected, there is a need to explain discrepancies, which leads to learning. Many things can be learned: a new case, a new knowledge structure, new knowledge learned through explanation, a new way to index, and so on. In the most successful automated case-based reasoners, learning is attempted even when knowledge is incomplete.

But CBR is more than a set of procedures that carry out analogical reasoning using near analogs. It suggests a cognitive architecture that unifies understanding and problem solving and that integrates reasoning (for understanding, problem solving, and explanation) with memory and learning. Analogical reasoning is central; thus, the major processing cycle in this architecture is "recognize and apply. " The reasoner is constantly matching its new goals and situations with what is in memory to find the most usefully similar, already experienced situations. When it finds a useful match, it tries to apply the suggestions made by that experience to the new situation. The reasoner collects up the results of its actions, analyzes them, and updates its memory accordingly. Learning is a natural consequence of applying knowledge to new situations, noticing and analyzing the results, and inserting the experience into memory. Key inference processes are those that allow instantiation and adaptation of old structures to fit new situations: adaptation, extraction, merging, comparing and contrasting, and so on. Inference is also done in the service of retrieval and case-adaptation procedures, for example, to reinterpret a situation to allow recognition of something in memory that can be used to process it. And analogical reasoning is always done with a purpose.

CBR attempts to unify and explain the processing behind human memory, reasoning, learning, and performance. Implementation of its algorithms in artificial intelligence problem-solving systems (Kolodner, 1993) has helped explicate the variety of computational processes needed for successful recognition and reuse of experiences. As such, the approach provides us with a number of interesting implications for helping people learn to be good reasoners.

If use of cases is central to CBR, then clearly a major issue for CBR is the indexing problem: identifying old situations that are relevant to a new one. A reasoner cannot apply an old situation to a new one until after the old situation is recalled. Conceptually, this is the problem of accessing appropriate knowledge at appropriate times. Specifically, it means recognizing the applicability of an old situation to a new one. This is not a new concept for psychology. The phenomenon of encoding specificity deals with the knowledge-access problem, telling us that the reasoner is more likely to recall an old situation to the extent that the description of the new situation matches the encoded description of an old one. CBR proposes a computational model of the processing that allows this kind of recognition.

In short, three sets of procedures are involved: (a) those that operate when cases or experiences are encoded and stored in long-term memory, (b) those that operate at retrieval time, and (c) partial matching procedures that mediate between the two. At insertion time (encoding), a reasoner interprets a situation, at least to some extent, and identifies at least some of the lessons that can be taught and when those lessons might most productively be applied. The case is labeled, or encoded, according to its applicability conditions, in other words, the circumstances in which it ought to be retrieved. The most discriminating labels on a case will be derived by a reasoner who has taken the time and effort and that has the background knowledge to carefully analyze a case's potential applicability. A novice, or a reasoner who has not done a careful analysis of applicability, may not be able to assign discriminating labels or may assign these labels only on the basis of superficial features of the case (e. g. , Gick & Holyoak, 1980; Lancaster & Kolodner, 1987; Ross, 1986).

At retrieval (recognition) time, retrieval processes use aspects of the description of a new situation as a probe into memory, looking for cases labeled in such a way that they match the probe. The extent to which a reasoner is willing or able to interpret the new situation and the quality of the representation of the new situation that ensues determine the quality of the probe into memory. An uninterpreted situation is likely to yield poorer access to the contents of memory than one that is more embellished. The more creative a reasoner is at interpreting a situation, the more likely that person is to find relevant knowledge and experience to use in reasoning about it. Also important to retrieval are partial matching procedures, procedures that have the capability of reconciling the differences between closely but not exactly matching details.

Given this set of processes, the extent to which a reasoner is good at identifying relevant old situations depends on several things: (a) how well and how completely the reasoner interpreted the old situation (and the trace left of that interpretation), (b) how well and how completely the reasoner can interpret the new situation to identify what is important about it, and (c) how good the reasoner's partial matching procedures are. Thus, there is an interesting and important interplay between encoding procedures and retrieval procedures. Excellent encoding procedures that foresee a large range of the possible situations in which a case might be useful make the job of situation assessment procedures easy. Excellent situation assessment procedures that are able to reinterpret a new situation in a way that might match something that happened previously allow encoding procedures to be weaker. But excellent procedures of both kinds require a tremendous amount of knowledge, which is also unevenly accessible (Barsalou, 1987). Thus, the best recall will be generated by reasoners who do the best job they can with whatever knowledge is available at both encoding and retrieval times.

The extent to which an old situation is interpreted also determines the extent to which it can be useful in guiding reasoning: With only a description of a problem situation and its solution, an old solution can be repeated. If, in addition, the reasoner knows whether the solution succeeded when applied, there is a basis for deciding whether or not to reuse an old situation. If, in addition, the reasoner knows what happened as a result of applying the solution, more reasoned judgments are possible (e. g. , "Does that result make sense in my new situation?"). If, in addition, factors responsible for the result (success or failure) are known, even more reasoned judgments are possible (e. g. , "Are the aspects responsible for failure in the old situation present in the new one? If not, the same failure might not recur. ").

According to CBR, performance is enhanced by one's ability to draw from a rich experience base. Learning occurs when the individual extends this knowledge by continually and thoughtfully incorporating new experiences. CBR suggests that providing a learner with opportunities for interpreting a new situation or problem and identifying significant descriptive, explanatory, and predictive features of a problem and its solutions will facilitate encoding new situations in reusable ways and promote reminding of already experienced situations.

CBR gives failure a central role in promoting learning. When a reasoner's expectations fail, it is alerted that its knowledge or reasoning is deficient, and a need to learn arises. Similarly, an unsuccessful outcome or solution warns the reasoner of a deficiency and, therefore, a need to learn. Failure at applying an old case in a new situation triggers explanation that might result in reinterpreting (reindexing) old situations or discovering new kinds of interpretations (indexes). Useful feedback from the world is critical to interpreting failure. A reasoner that is connected to the world will be able to evaluate its solutions with respect to what results from them, allowing indexing that discriminates usability of old cases and allowing good judgments later about reuse.

We can summarize what CBR suggests about learning in six statements:

• A case-based reasoner learns by acquiring cases and indexing them; thus, having experiences that can contribute to learning is crucial.

• Failure at applying an old case in a new situation triggers explanation that might result in reinterpreting (reindexing) old situations or discovering new kinds of interpretations (indexes).

• A reasoner that is connected to the world will be able to evaluate its solutions with respect to what results from them, allowing indexing that discriminates usability of old cases and allowing good judgments later about reuse.

• Failure is a good motivator of explanation, focusing a reasoner on what it needs to learn.

• As the reasoner's knowledge and experience grow, it will be better able to index its cases.

• The better the analysis of an experience, the better the indexes, the better the later access, and the better the learning.

As a plausible model of cognition, CBR provides several suggestions about educating. Most important, it suggests a style of education in which students learn by having rich experiences that motivate the need to learn and that give them a chance to apply what they are learning. In such an environment, students might engage in solving a series of real-world problems (e. g. , cleaning up the Boston Harbor, generating an emergency plan for the next heat wave, designing locker organizers, routing taxis or airplanes, or Middle East peace negotiations), either for real or through realistic simulation. Participation in design and problem-solving activity, especially when students must make something work, gives them the opportunity to notice what they need to learn, experience the application of that knowledge, and learn how it is used. Designing locker organizers, for example, requires the students to understand the variety of ways lockers are used; concepts of geometry; and concepts about physical structures, supports, and materials. They might engage in taking surveys and learn both math concepts (e. g. , sampling, averaging, and probabilities) and social sciences concepts (e. g. , question asking). They might learn concepts in geometry through drawing and manipulation of shapes. They might learn physics concepts from consideration of the kinds of support structures their locker organizers need, and so on.

CBR makes several other suggestions—on the basis of its focus on the role of failure, its focus on indexing as the key to reuse, and its focus on the role of experience in reasoning—about the kinds of experiences that encourage good learning and how to sequence those activities in a curriculum, the kinds of materials and resources that ought to be made available to students engaging in these experiences, how to promote success and learning as students engage in these activities, how to promote transfer and facilitate cognitive flexibility, and software tools that encourage learning. ²

Modern approaches to education suggest that student learning experiences should resonate with their experiences outside the classroom so as to engage and motivate the children and give them a way to get started. Problem-based approaches to education (Barrows, 1985; Barrows & Kelson, 1995) add that students should engage in problem-solving enterprises that are purposely complex, ill-structured, and open-ended, lending themselves to several interpretations or solutions and painting a cohesive, holistic view of an issue or situation, so that students learn skills and facts in situations of realistic complexity. CBR's focus on the role of failure in learning suggests the importance of acquiring feedback on decisions made, to be able to identify holes in one's knowledge and to generate goals for additional learning. CBR's approach emphasizes the need for students to carry out and test their ideas, not only think about them. The most effective problems for learning, it suggests, will be those in which students get feedback along the way that allows them to recognize the holes and misconceptions in their knowledge, refine their knowledge and reasoning strategies, and evaluate the goodness of their knowledge and reasoning strategies. CBR further suggests that problems should present difficulties for students to give them an opportunity to see where the real complexity in situations or domains lies and that success at solving every problem is not necessary, as long as such affordances for learning are present.

Solving real-world problems is much more difficult for both students and teachers than the kinds of activities that students typically do in today's schools. It is essential, therefore, to provide "scaffolding" (Blumenfeld et al. , 1991; Brown, 1981; Vygotsky, 1978), or help, for students as they are solving problems to allow them to be successful. CBR suggests that libraries of relevant cases be made available to students as they are solving problems (Kolodner et al. , 1996; Narayanan & Kolodner, 1995; Schank et al. , 1993). Such cases, even though they have been experienced by others, can suggest issues to focus on and solutions to problems, warn of potential pitfalls, support projection of the effects of a chosen solution, and so on, facilitating solution of more complex problems than students could solve by themselves. Examination of the locker organizers on the market, for example, might help students identify important issues to consider, might give students who are designing locker organizers ideas that can be incorporated into new designs, and might provide them with a way to predict the results of designs they generate. Examination of survey instruments and the pros and cons of each would help them in similar ways to design their locker-use surveys.

The CBR community has designed case-based design aids (Domeshek, Kolodner, & Zimring, 1994) and other kinds of case libraries (Bell, Bareiss, & Beckwith, 1993; Kass, Burke, Blevis, & Williamson, 1993) that store the cases or experiences of others (often experts) for problem solvers to peruse while reasoning. Cases in the case library should include cases that can help students identify issues that need to be addressed (learning issues), that suggest potential solutions or parts of solutions or ways of addressing issues, and that can help with projecting effects of potential solutions. They should be indexed in ways that resonate with the issues that students are uncovering as they solve a problem. Students should also be encouraged to reuse their own experiences and helped to learn how to recognize if some case or experience might be useful.

Also necessary for successful problem solving is that students know how to address problems and navigate their complexities. Problem-based learning suggests the use of well-structured white boards for keeping track of facts, ideas, and what needs to be learned. Reciprocal teaching (Brown & Palincsar, 1987) suggests that teachers, acting as coaches, model reasoning processes for students in ways that encourage students to adopt those reasoning strategies for themselves. CBR suggests that teachers should help students learn to carry out CBR well. The analogy literature suggests that students need help making analogies—whether to cases in a case library, to examples in a text, or to their own experiences—especially help with mapping. A major difficulty in using cases to reason lies in realizing that suggestions made by cases must be verified. Other difficulties are in recognizing what to focus on, carrying out adaptations, and determining which of several recalled cases might be most applicable. CBR can be learned, researchers believe, by having students carry CBR out and then drawing attention to and reflecting on its use. Several kinds of software tools might also be helpful in aiding the use of analogous cases: tools for lining up cases or examples next to each other so as to identify their matching pieces (mapping), tools for generating and testing ideas derived from cases, tools for comparing and contrasting cases, tools for articulating what is important, and tools for carrying out or simulating adaptations.

Experience, by itself, does not ensure learning. We recognize that someone has learned from an experience when he or she can apply what is learned in a new situation (transfer). Recall that CBR tells us that application of knowledge learned in one situation to another situation depends on both (a) the extent to which the old situation was interpreted and its lessons learned were articulated and recorded and (b) the accessibility of the old situation when the new one is encountered. CBR suggests that we can promote transfer by helping students to extract and clearly articulate what they have learned, helping them make those articulations rich in the right ways, and helping them index their experiences well (by predicting the circumstances in which an experience's lessons might be appropriately applied).

The transfer literature (e. g. , Salomon & Perkins, 1989) tells us that transfer requires looking forward at the time of an experience or at the time knowledge is acquired to consider the circumstances in which it might prove useful and looking backward when solving a new problem to consider whether such an experience has been encountered before. CBR tells us much the same thing, but more concretely. Looking forward, CBR suggests, should include identifying what lessons an experience teaches and predicting the circumstances when those lessons might be appropriately applied (Kolodner, 1993). Looking backward involves incrementally making plausible elaborations on a new situation, attempting to reassess the situation from a different point of view and rerepresent it when such elaboration is not fruitful, and using more complicated search strategies to search for related types of situations when that is not fruitful. Experience from building case libraries suggests a way of promoting these kinds of reflection with software: We might ask students to create and use each other's cases. They must look forward, thinking about applicability, as they create cases for others to use that articulate what they have learned from a situation. In thinking about how they might query a library of cases, they are forced to carry out looking-backward procedures.

CBR also makes suggestions about promoting cognitive flexibility. Cognitive flexibility refers to knowing a concept in its full complexity, so as to be able to effectively use it in a variety of situations. Because a single experience with a concept shows only one way it can be used, cognitive flexibility theory (Spiro, Coulsen, Feltovich, & Anderson, 1988) suggests that concepts be revisited from several points of view. CBR suggests that knowledge will be more accessible, flexible, deeply learned, and accurate if learners have the opportunity to encounter (firsthand or by report) multiple situations in which the knowledge is used and multiple ways in which similar situations are addressed and if students have the opportunity to reuse and try out knowledge gained through experiences. Making case libraries available to learners not only can help them generate ideas and solutions, it can also promote flexibility. CBR also suggests that problems be written in such a way that they promote reminding of earlier problems in which the same concept was used and that teachers help students notice that concepts are being reused and abstract from the range of ways in which the concept is encountered.

It is important for any curriculum to cover its learning outcomes. For purposes of developing an authentic understanding of concepts, flexibility, and transfer, problems that form a curriculum should be chosen so that key concepts are visited over a number of cases. Providing coverage in a problem-solving domain, according to CBR (Kolodner, 1993), means that students should encounter typical problems, typical solutions, typical situations in which each problem and solution occur, typical reasoning pitfalls, and several important atypical situations.

On the basis of these analyses, experiments with educational practice and educational software are currently being carried out in two major centers: the Institute for the Learning Sciences (ILS) at Northwestern University, headed by Roger Schank, and the EduTech Institute at Georgia Institute of Technology, headed by Janet Kolodner. In all of these endeavors, students learn by carrying out a mission that requires the learning of some critical knowledge, set of complex skills, or both. In all, the computer environment supports their success at problem solving. All include case libraries that make expert knowledge available to students as they are reasoning. Some case libraries are in the form of video (e. g. , ASK systems [Ferguson, Bareiss, Birnbaum, & Osgood, 1992]), some in the form of text and graphics (e. g. , case-based design aids [ Narayanan & Kolodner, 1995; Zimring, Do, Domeshek, & Kolodner, 1995], and World Wide Web pages), and sometimes classes are orchestrated so that the experts themselves are available with their stories as students need advice.

Schank's group at ILS is creating computer environments, called Goal-Based Scenarios (Schank et al. , 1993), to support self-paced learning of some piece of a curriculum, with a focus on informal education and training. Each Goal-Based Scenario simulates the real world and includes all of the content, tools, workspaces, and simulated experts that students need to carry out their mission. Broadcast News (Schank et al. , 1993), Sickle-Cell Counselor (Bell et al. , 1993), and YELLO (Kass et al. , 1993) are the most highly publicized of the Goal-Based Scenarios. Using Broadcast News, high school students create a newscast of some historical event. To create it, they must look up and verify facts, determine how best to describe the characters and events, and determine to what extent they should relate these events to other world events. As they are working, the system analyzes their work and makes available to students the advice of historians, newscasters, and journalists, all captured in stories and on video. Historians might comment on the importance of an event, journalists and newscasters on how to explain things well to an audience, and so on. Using the Sickle-Cell Counselor, visitors to Chicago's Museum of Science and Industry learn about genetics and sickle-cell anemia by pretending they are genetic counselors. They counsel the software's simulated couples about whether or not to have children. They must go through the actions of taking blood, analyzing it, considering the indicators in both parents, and plotting the probability of whether the children will have sickle-cell anemia. The program helps them to walk through the actions by providing a simulated needle for taking blood, simulated machines for analyzing blood, and so on. The program makes available knowledge about sickle cell, genetics, processing blood, computing probabilities, and so on. Users ask to see resources to learn what they need to successfully counsel their simulated couple. Video clips of a technician, doctor, and counselor are made available when the program determines that users need answers to questions or when users ask. YELLO is used to train salespeople in sales of Yellow Pages ads. As in the Sickle-Cell Counselor, the user interacts with simulated customers.

In each of these systems, the computer provides a complete environment for carrying out the designated mission—a simulated world to work in, appropriate factual knowledge, ways of trying things out, and advice. Goal-Based Scenarios deliver a course, so to speak, to the individual students who use them. Guidelines for building Goal-Based Scenarios have been drafted (Schank et al. , 1993) so that a broad range of these software environments can be built. These guidelines are being used as well to plan training courses independent of computer use. In several training courses, students are given a mission to accomplish that has them solving workplace problems and using personnel in the company to simulate the environment in which the problem solving would take place were it real. In this way, new employees can learn workplace skills in an authentic way, but problems can be chosen to orchestrate learning well, and students do not have the pressure of making real company decisions with real consequences.

Kolodner's group at EduTech is working with teachers to develop classroom activities for science and technology education. In their approach, called Learning By Design (Gertzman & Kolodner, 1996; Kolodner et al. , 1996; Narayanan et al. , 1995), focus is on the use of engineering design problems to promote learning. The lung, Appalachian Trail, arthropod robot, and locker-organizer problems are being used at the middle school level, and kite design and kiosk problems are being used for undergraduate design education. Attention has been given to designing four key components of the learning environment: curriculum materials and sequencing, a software environment that promotes successful problem solving, case libraries and other resource materials, and guidelines to help teachers promote successful problem solving and learning. CBR has been wedded to problem-based learning (Barrows, 1985), an educational methodology that provides principles of practice to go along with CBR's conceptual principles. Problem-based learning, like CBR, recognizes the importance of learning from work on real-world problems and the need for reflection of the right sorts to cement learning. To deal with helping students be successful, it promotes work in collaborative groups and training teachers as facilitators.

An attempt is being made in the middle school project to cover large parts of the middle school science curriculum using a sequence of design problems that students solve. Students learn about body systems from the lung problem, nutrition from the Appalachian Trail problem, and so on. A unit on the human body might include both of these problems plus others. Software development focuses on issues that come up across problems, for example, keeping good records of deliberations, comparing and contrasting cases, deciding between alternatives, and deciding what research to do. Software aims to play the teacher's facilitator role for small groups of students, because the teacher cannot be available to help every group all the time. But it does not take the place of the teacher completely; the teacher is still in charge of facilitating large-group discussions that promote reflections of the right kind. Software already piloted in classrooms includes McBagel (Guzdial et al. , 1996; Narayanan et al. , 1995), which helps students keep track of their deliberations when they are working together, helping them at the same time to discover the kinds of records they need to keep to be successful in solving complex problems, and WebCaMILE (Guzdial et al. , 1996), which helps students collaborate with each other when they are working as individuals. To this set, we hope to soon add software that helps students carry out design reasoning well, promotes deliberations about which of several alternative solutions is best, helps students apply cases well, helps them plan their time and collaborative efforts, helps them learn how to access and use resources well, and promotes the right kind of reflections on their experiences. Case libraries support some of the problems, and we are developing authoring tools for case libraries, both for curriculum developers or teachers to use as they are putting together resources to support new problems and for students to use to record their experiences for other students to learn from. Appropriate simulations and other off-the-shelf tools are made available in the software environment as appropriate for each problem. The use of such a software environment across many different problem-solving and design experiences, researchers predict, will promote learning of a full range of cognitive, social, and self-directed learning skills; the ability to apply them flexibly in many kinds of situations; and the promotion of deep learning of standard curriculum content.

Does any of this work? It is early, but the indicators are positive (e. g. , Gertzman & Kolodner, 1996). Teachers and trainers who use CBR-informed materials come back energized. Teachers feel that they are able to reach more of their students with this methodology. Both those at the top and those at the bottom are drawn in more by these activities than they are in a normal "aim-toward-the-middle" classroom. Concepts and skills are being learned, teachers believe, in ways that will encourage students to remember and reuse them. Students surprise the teachers with ideas they come up with and the connections they are able to draw. Initial data show that, indeed, students are learning and often better than students in a traditional classroom.

At the same time, some teachers and students are somewhat overwhelmed by the new way of doing things. To effectively learn from complex problem solving and design experiences, students need to learn to be good problem solvers, designers, planners, reflective practitioners, collaborators, and communicators, in addition to learning the curriculum content. It is clear that one-shot approaches do not work; teachers need to hone their skills at working in a problem-solving environment, and students need to incrementally learn collaborative communication, design, and other skills over the course of solving and learning from several problems. Because these endeavors are new, analyses so far, both at Northwestern University and Georgia Institute of Technology, have necessarily been analyses of situations in which teachers and students are engaged in these activities and in using our software for the first time. These analyses can give us ideas about how to proceed in developing software, guidelines for teacher practice, guidelines for problem creation, and so on. When students are learning from collaborative problem solving and design activity on a regular basis and when they are making regular use of the software environments being created, we will know for sure the real impact of CBR's suggestions.

This example was written by Hari Narayanan (personal communication, 1995) based on examples (some real and some hypothetical) presented by Domeshek and Kolodner (in press) in their work examining the role cases play in design and in their design of Archie (Domeshek & Kolodner, 1992), a system for aiding architects as they do design, and DesignMUSE (Domeshek et al. , 1994), an authoring system for construction of case-based design aids. It is quite similar to other more complex real examples of architects in action.

CBR's suggestions are consistent with constructivist philosophies and practices, especially problem-based theories of practice (e. g. , problem-based learning [Barrows, 1985; Barrows & Kelson, 1995] and anchored instruction [Cognition and Technology Group at Vanderbilt, 1993]), and it makes several of the same observations and recommendations as other cognitive theories of learning (e. g. , cognitive flexibility theory [ Spiro et al. , 1988]). The level of detail that comes from CBR's computational models allows it to explain why some of the exemplary practices of teachers work, to add specificity to some practices, and to add guidelines and software tools to augment current practice.