Case-based reasoning (CBR) systems rely on structured knowledge called cases for reasoning. These cases typically represent examples or prior experiences from a task domain. Acquiring adaptation-relevant knowledge in case-based reasoning systems has proven to be a challenging problem. Such knowledge is typically elicited from domain experts or extracted from the case-base itself. The first approach is often limited by practical concerns, such as time and cost, while the second is limited by the knowledge present in the case base. Fortunately the Web contains a vast source of knowledge on a variety of topics, compiled by experts and volunteers, which can be navigated by artificial agents. This research focuses on the problem of acquiring case adaptation knowledge from Web-based resources. The primary hypothesis of this research is that task-relevant knowledge for a CBR system can be mined from Web-based resources on demand. This dissertation proposes integrating the knowledge discovery process with the traditional case-based reasoning cycle. Knowledge discovery from Web-based resources is then applied to the problem of acquiring case adaptation knowledge for CBR systems. Because the web contains knowledge covering a variety of domains, a generic knowledge discovery framework is required in order to discover and reason about task-relevant knowledge. A generic knowledge discovery framework for case adaptation knowledge is implemented in the system WebAdapt. To increase its generality, WebAdapt relies on minimal pre-coded domain knowledge and acquires all other task-relevant knowledge on-demand. This dissertation discusses WebAdapt's basic model and the necessary components for a generic knowledge discovery framework for CBR, including (1) automatic detection and recovery from failures in the search process, (2) resource selection, and (3) knowledge retention. A set of empirical results are then presented to evaluate the efficacy of Web mining for case based reasoning. The contributions of each component in the framework are also judged based on empirical results.

Case-based problem-solving systems reason by retrieving relevant prior cases and adapting their solutions to fit new circumstances. The ability of case-based reasoning (CBR) to reason from ungeneralized episodes can benefit knowledge acquisition, but acquiring the needed case adaptation knowledge has proven challenging. This paper presents a method for alleviating this problem with justin-time gathering of case adaptation knowledge, based on introspective reasoning and mining of Web knowledge sources. The approach combines knowledge planning with introspective reasoning to guide recovery from case adaptation failures and reinforcement learning to guide selection of knowledge sources. The failure recovery and knowledge source selection methods have been tested in three highly different domains with encouraging results. The paper closes with a discussion of limitations and future steps.

If given manually-crafted goal selection knowledge, goal reasoning agents can dynamically determine which goals they should achieve in complex environments. These agents should instead learn goal selection knowledge through expert interaction. We describe T-ARTUE, a goal reasoning agent that performs case-based active and interactive learning to discover goal selection knowledge. We also report tests of its performance in a complex environment. We found that, under some conditions, T-ARTUE can quickly learn goal selection knowledge.

Our poster presentation illustrates how to include wrong answers in test analyses using Response Spectrum Evaluation (RSE) procedures to track answer patterns on an answer-by-answer basis. RSE is a statistical procedure adapted from the multinomial [3] that bypasses the linear dependency problem so that alternative (wrong) answers can be included in data-mining analyses. Thus, the study of the dynamics of learning events can be conducted on an answer-by-answer basis. Previous investigations [3] using this procedure have revealed: 1. The selection of answers is the result of the way students interpret the test questions. 2. These interpretations are directly inferable from the answers selected (or presented). 3. Selection procedures involve a number of strategies that are characteristic of each student, providing diagnostic information that can inform teaching. 4. This information is of more value to teachers who focus upon teaching how to think and how to learn instead of reproducing course content. 5. Some students show systematic development similar to the sequence described by the clinical observations of Piaget [1], while others show deterioration in the reverse direction. 6. Some students systematically shift from the right answer on the easy questions to particular types of “wrong” answer when their ability breaks out of “all” or “nothing” thinking (without considering other options) into more intellectually flexible mindsets. RSE is the only procedure with this detection capability. 7. The focus upon the right answers in the psychology of test-taking reinforces closedminded thinking on the part of students taking the test, meaning that if the objective of teaching is profound understanding, the focus upon “right” answers is psychologically invalid. 8. The dynamics of learning revealed by the RSE procedures are non-linear and multichotomous, meaning that the use of total-correct scores to assess student

Acquiring adaptation knowledge for case-based reasoning systems is a challenging problem. Such knowledge is typically elicited from domain experts or extracted from the case-base itself. However, the ability to acquire expert knowledge is limited by expert availability or cost, and the ability to acquire knowledge from the case base is limited by the the set of cases already encountered. The WebAdapt system [20] applies an alternative approach to acquiring case knowledge, using a knowledge planning process to mine it as needed from Web sources. This paper presents two extensions to WebAdapt's approach, aimed at increasing the method's generality and ease of application to new domains. The first extension applies introspective reasoning to guide recovery from adaptation failures. The second extension applies reinforcement learning to the problem of selecting knowledge sources to mine, in order to manage the exploration/exploitation tradeoff for system knowledge. The benefits and generality of these extensions are assessed in evaluations applying them in three highly different domains, with encouraging results.

How to endow case-based reasoning systems with effective case adaptation capabilities is a classic problem. A significant impediment to developing automated adaptation procedures is the difficulty of acquiring the required knowledge. Initial work on WebAdapt [1] proposed addressing this problem with “just-in-time” knowledge mining from Web sources. This paper addresses two key questions building on that work. First, to develop flexible, general and extensible procedures for gathering adaptation-relevant knowledge from the Web, it proposes a knowledge planning [2] approach in which a planner takes explicit knowledge goals as input and generates a plan for satisfying them from a set of general operators. Second, to focus selection of candidate adaptations from the potentially enormous space of possibilities, it proposes personalizing adaptations based on learned information about user preferences. Evaluations of the system are encouraging for the use of knowledge planning and learned preference information to improve adaptation performance.

Making case adaptation practical is a longstanding challenge for case-based reasoning. One of the impediments to widespread use of automated case adaptation is the adaptation knowledge bottleneck: the adaptation process may require extensive domain knowledge, which may be difficult or expensive for system developers to provide. This paper advances a new approach to addressing this problem, proposing that systems mine their adaptation knowledge as needed from pre-existing large-scale knowledge sources available on the World Wide Web. The paper begins by discussing the case adaptation problem, opportunities for adaptation knowledge mining, and issues for applying the approach. It then presents an initial illustration of the method in a case study of the testbed system WebAdapt. WebAdapt applies the approach in the travel planning domain, using OpenCyc, Wikipedia, and the Geonames GIS database as knowledge sources for generating substitutions. Experimental results suggest the promise of the approach, especially when information from multiple sources is combined.