Learning to Adapt Cases to New Situations

Our approach to case adaptation learning models a transition from general, non-operational adaptation knowledge to more specific and operational knowledge. The system learns to improve its adaptation capabilities by case-based reasoning applied to the case adaptation process: a trace of the solution process from an adaptation problem is saved to be replayed when similar adaptation problems arise in the future [Leake1995a,Leake et al. 1995]. Initially, the system knows about a small set of abstract transformation rules and memory search methods, but by acquiring new cases describing specific adaptations, DIAL expands on this knowledge with additional adaptation knowledge.

DIAL's adaptation component takes two inputs: an instantiated disaster response plan and a description of the problems in the response plan that must be repaired. When presented with an adaptation problem, DIAL's adaptation component performs the following steps:

1. Case-based adaptation: DIAL attempts to retrieve a similar adaptation case to be reapplied directly; if successful, a candidate solution is passed to step 3.

2. Rule-based adaptation: When no relevant prior case is found, DIAL selects a transformation associated with the problem (e.g., substitute a new plan step). A knowledge goal is generated to find the information needed to operationalize the transformation, and it is passed to a planning component. The planner uses introspective reasoning about memory search strategies [Leake1995b] to guide the search for the needed information. If the information is found, it is used to apply the selected transformation to the retrieved response plan, and a new adaptation case is stored. If the information is not found, the process continues with step 4, manual adaptation.

3. Plan evaluation: The result is evaluated for compatibility with explicit constraints from the response plan, with a human user performing backup evaluation.

4. Manual adaptation: If autonomous case adaptation fails, an interface allows a user to guide the adaptation process, and the results may be stored as a new case.

5. Storage: When DIAL successfully adapts a response plan, it learns by storing (1) the new response plan case, (2) memory search cases encapsulating the memory search steps performed during case adaptation, and (3) adaptation cases, which encapsulate information about the adaptation problem and its solution.

The basic principles of the adaptation process are shown by the example of developing a response plan for a 1994 flood in Allakaket, Alaska. When DIAL receives the story about Allakaket, it retrieves the most similar disaster case in memory, a flood in Bainbridge, Georgia. Part of the response to that flood was to build walls of sand bags to protect the area from water damage as the flood waters rose. In Bainbridge, volunteers helped build sand walls; DIAL generates a knowledge goal to find people who could fill the same role in Allakaket. A previous manual adaptation episode suggests finding able-bodied local residents. However, most of the able-bodied people in Allakaket were unavailable because they were helping to fight fires in the northwest. This prompts a new problem, that the desired role-fillers are unavailable. DIAL has no similar adaptation cases, so it falls back on rule-based memory search to attempt to find a substitution. It checks constraints on possible role-fillers and finds that the volunteers were under the authority of the police. Searching for others under police authority, it finds local prisoners. Prisoners are suggested to build the flood walls, and, when evaluated as reasonable by a human user, the transformation is performed and saved. This new case can be used as the basis for performing future adaptations.

• Effects of learning response plan cases and adaptation cases