Case-Based Reasoning for Distributed-Learning Environments

---

CBR Overview

Case-based reasoning is one of the major reasoning paradigms in artificial intelligence. A CBR reasoner^[1] solves new problems by retrieving a similar problem from a case base, which stored the experienced solutions to past problems. When the reasoner cannot find a solution that is similar enough to solve the new problem, CBR will adapt the solution of a relatively similar problem to the new one. In principle, CBR is different from other AI reasoning approaches such as rule-based reasoning (Yang et al., 2003) and model-based reasoning. In the case of the rule-based reasoning system, for example, expert systems, the rules reflect the certain relationships between the problems and their solutions. The rules can be designed from the text-based documents or from domain expert’s experience and know-how. However, the cases in CBR systems reflect or document the relation between the problems and their solutions, which were obtained from past experience. Such a relationship might not be certain, and the solution for the similar problem may not be unique. Therefore, the CBR system has to be able to learn a new solution for the same problem. Somehow, CBR can provide an alternative to rule-based reasoning systems, and it is suitable for application domains in which the theory is very weak or incomplete.

All in all, CBR is a methodology for solving problems in the real world. In general, CBR involves the following main steps for new problems:

• Retrieving similar cases from the case bases, which were created from historic experience

• Selecting a best case from the retrieved cases

• Adapting a similar solution to the new problem if necessary

• Applying the solution to the problem

• Evaluating the solution based on the problem-solving result

• Creating a new case if the case does not exist in the current case bases

Aamodt and Plaza (1994) described such reasoning steps as a typical cyclical process: Retrieve, Reuse, Revise, and Retain, which is called “the four Rs”:

1. RETRIVE the most similar cases from previously experienced cases

2. REUSE the information and knowledge in the retrieved cases to solve the new problem

3. REVISE the solution, if necessary, by adapting the similar solution

4. RETAIN the new solution as a new case for an existing case base

From the viewpoint of problem-solving methods in CBR systems, CBR reasoning could be classified into two main types: conversational case-based reasoning (CCBR) (Aha et al., 1999) and automated case-based reasoning (ACBR). CCBR is used to solve the problems that may not be entirely clear at the beginning of the reasoning process. It requires users to initially supply a brief partial description of the problem. With ongoing reasoning, CCBR supports interactive problem assessment to construct a query for the problem, using well-designed questions. In terms of the answer obtained from users, CCBR keeps reasoning until reaching the final solution. CCBR is suitable for application domains, where an entire description of the problem is difficult, and where the solution may be dynamic for the problem, such as in task planning.

ACBR is a traditional type of CBR. It requires users to determine the problem relevance of each attribute for the problem, and to have detained domain knowledge. Once the problem is described entirely with the detailed information for each attribute, ACBR automatically performs the reasoning process to find the solution for the given problem. This is also a noninteractive reasoning process between the systems and users. ACBR is more useful in those application domains where the environment is relatively static, such as in maintenance and diagnosis.

Today, CBR has been widely utilized for various application domains, such as medicine, diagnosis, knowledge acquisition, help-desk, design, planning, and maintenance. It has also been applied to learning and teaching for intelligent support.

Applying CBR to MAS-based DLEs

As an effective problem-solving methodology, CBR can be used to support learning and teaching, and to perform coordination and collaboration for the agents in MAS. Following are the descriptions of these applications.

Case-Based Learning

The research results in learning science show us that deep and effective learning is best promoted by situating learning in authentic activity. Anchored instruction, project-based learning, problem-based learning, and other constructive approaches to classroom practice all focus on putting students into situations where they must make hypotheses, collect data, and determine which data to use in the process of solving a problem or participating in some kind of realistic investigation (Kolodner et al., 1996). Such learning activities can be improved by using the CBR method, because it facilitates students’ own case acquisition by presenting students with information about their experiences, in the form of relevant cases, when cases are likely to be useful. For example, Kolodar et al. (1996) proposed to use a CBR approach to refine the definition for problems in the problem-based learning.

Case-Based Coordination for Agents in MAS

As mentioned above, DLEs are developed with MAS technology, supported by broadband communication networks. MAS is one of the key technologies. In the DLEs, it is clear that the agent in MAS must have coordinating ability to interact the dynamic environments for providing intelligent support. A MAS- based DLE is a work team that consists of all of the agents. To coordinate these agents into a team effectively in dynamic environments, we need a good multiagents infrastructure and an effective means to determine the role of each agent and maintain and achieve the full team goal. Giamppapa and Sycara (2001) proposed to use a CCBR reasoner as a task planer for coordinating agents as a task team in MAS. In this section, we discuss applications of the CCBR reasoner, NaCoDAE (Navy Conversational Decision Aiding Environment), to DLEs for coordinating and collaborating agents in MAS.

NaCoDAE Overview

NaCoDAE (Aha et al., 1998) is a CCBR system, which was developed at Naval Research Laboratory, USA. It is a CBR tool oriented to planning applications. NaCoDAE helps users to solve problems by interacting with the system. The conversational process begins with the users providing an initial partial description of the problem. NaCoDAE responds to the partial problem by providing a ranked solution list. The users interact with the list, either refining their problem description by answering the selected questions or directly selecting the solution to apply to the problem. In NaCoDAE’s case bases, each case C has three components:

• Summary: Brief text Ctthat partially describes C

• State: A set of Cqa <question, answer> pairs

• Solution: A sequence Cs of actions for responding to state Cqa

The summary and state are used to construct a problem description for caseC. Given the free-text P for the problem description, NaCoDAE computes the similarity score for all cases in the case bases. Based on the scores, NaCoDAE identifies, ranks, and displays a set of the cases that are most similar to the problem. At the same time, it also displays the set of questions that are the most frequently unanswered ones in this set of the cases. Through such interactive actions between the users and NaCoDAE, a final solution will be applied to the problem. Figure 1 shows the interactive processes between the users and NaCoDAE. NaCoDAE has three main features that make it suitable for coordinating agents in MAS. First, it can work starting with a partial problem description. This is very powerful in the case of dynamic learning environments, where the problem may not be fully understood at the beginning of the problem- solving process. Second, NaCoDAE can cautiously revise its list of similar cases as long as detailed information is provided to the system by either an agent or the user. Last, the cases can be edited to store the free-text data, including agent capabilities and queries.

Figure 1: NaCoDAE: A CCBR system Source: Adapted from Aha and Maney (1997).

Applying CCBR to Coordination of Agents in MSA

As mentioned in the previous section, the coordination of agents in MAS as a good agent team for DLEs in order to provide effective learning environments to the students and teachers is a vitally important issue. Specifically, the coordination in MAS consists of deciding the role for individual agents, defining agent behaviors that are consistent with their functional description, and forming an agent team for performing tasks in the dynamic leaning environments. There are several coordination techniques for coordinating agents in MAS. The most widely used techniques are capability-based coordination and team-oriented coordination. Capability-based coordination is a process, in which an agent dynamically discovers and interacts with other agents, based on its capability descriptions, and decides its role in MAS. Team-oriented coordination is a process that identifies and selects a group of agents to carry out with partial or complete description and to determine the role of each agent in the group by communicating with each other. To implement team-oriented coordination, the individual agent in MAS has to be a team-oriented agent that possesses three types of knowledge: the capabilities, the task requirements, and the social factors for role determination. In the DLEs, the capabilities could be an interface agent, information agent, teaching agent, learning object, and so on. The requirements could come from the student, professor, or administrator. The social factor to effect their role determination could be an authority to address task requirements, education institution structure, subordinate relationships, and so on. Considering that the requirements between agents might be partial descriptions for learning tasks, the CCBR is a good solution to realizing the team-oriented coordination in MAS. Figure 2 shows an infrastructure that incorporates with the CBBR reasoner, NaCoDAE, into MAS-based DLEs to perform team-oriented coordination.

Figure 2: CCBR-based coordinating infrastructure of MAS for DLEs

In the infrastructure, the NaCoDAE CCBR reasoner is embedded into a DLEs server agent for coordinating all the agents in MAS. Let us take a scenario to explain how the system works. The scenario is that Students A and B are going to work on a project as a team. The project requires knowledge from both Students A and B. When they log on to the Interface Agent 1 and Interface Agent 2 at different locations, the interface agents will send their requirements to the CCBR-based server agent. Based on their requirements for the project or problem description, which might be partial, the CCBR-based server agent will communicate with the task agents and information agents to decide which task agent can be responsible for the demands and which information agent can handle the necessary subject in terms of agent functional descriptions and their workloads. After the process of coordination is done, Task Agent 1 will become the two students’ teaching agent, and Information Agent 2 will take responsibility for providing the necessary learning object. Once a group of agents is formed as a team for a specific task, Students A and B can work together seamlessly on their project, from different locations and at different times.

^[1]A reasoning system is defined as a “reasoner” in this chapter. A case- based reasoning system is called a CBR reasoner.

---