Abstract
Traditional Case-Based Reasoning (CBR) systems often suffer from brittleness when the stage of a planning process in a new problem does not align with the stored case. This paper presents BRAINSTORMER, a system that utilizes model-based case adaptation to bridge these gaps. By integrating an explicit causal model of the planning process, BRAINSTORMER can transform abstract advice, such as proverbs, into operational data structures suitable for various planning stages, thereby enhancing the flexibility and scalability of knowledge-based systems.
Introduction
Knowledge-based systems frequently encounter "brittleness," where they fail if a problem does not precisely match pre-defined rules or cases. While Case-Based Reasoning (CBR) aims to mitigate this through adaptation, the adaptation process itself remains a significant bottleneck. A primary challenge is the mismatch in the stage of the planning process. For instance, a planner might receive advice relevant to plan refinement (e.g., selecting an actor) while it is currently at the plan retrieval stage. Standard CBR systems struggle to reconcile these differences. This research proposes a model-based approach to allow systems to flexibly utilize high-level, abstract advice regardless of the current planning step.
Methodology: The BRAINSTORMER System
The researchers developed BRAINSTORMER, a planner focused on counter-terrorism policy, to implement model-based case adaptation. The system employs a three-stage process to transform external advice into actionable internal structures.
- Translation to Canonical Form: External advice, often in the form of proverbs, is translated into a "canonical vocabulary." This vocabulary consists of standardized planning actions derived from a shared cultural model of planning.
- Causal Reasoning: This stage is the core of the model-based approach. BRAINSTORMER uses an explicit causal model of the planning process, represented as schemas, to reason about the advice. If advice pertains to a late-stage action, the system propagates these implications backward to earlier stages, such as plan selection, using the causal model.
- Operationalization: Finally, "if-added" rules convert these canonical representations into specific "operational" data structures that the planner can execute.
The system utilizes four primary inference mechanisms: schema activation, slot filling via abductive unification, redescription inference, and forward-chaining "if-added" rules.
Data and Knowledge Representation
As a symbolic AI project from 1992, BRAINSTORMER does not rely on modern large-scale datasets. Instead, its "data" includes:
- Proverbs as Abstract Cases: Proverbs like "an old poacher makes the best keeper" serve as culturally-shared cases. They are ideal because they encode generic strategies for complex planning and social scenarios.
- Domain Knowledge: A specialized knowledge base containing information on political and military policies related to counter-terrorism.
Key Insights
- Factoring Knowledge: A significant architectural contribution is the separation of knowledge about the planning process (the causal model) from the knowledge used to solve problems (the cases). This prevents redundant storage of planning logic across different cases.
- Flexibility through Explicit Models: By modeling its own planning steps, the system can "re-purpose" advice. It understands that recommending a specific actor type implies a specific plan type, allowing it to solve retrieval problems using refinement-level advice.
- Sublinear Scaling: The authors posit that because the causal model is general, the knowledge required for adaptation grows more slowly than the case base itself, improving efficiency as the system learns.
Results and Demonstration
The paper demonstrates BRAINSTORMER’s capabilities qualitatively. It shows how the system processes the proverb "an old poacher makes the best keeper," initially represented as an evaluation of an actor, and transforms it into a query for plan retrieval.
Through model-based reasoning, the system successfully bridges the "mismatch in stage," allowing the planner to retrieve a "prevent terrorism" plan even when the input advice was focused on actor selection. This proves that a model-based approach can effectively operationalize abstract advice for varying functional needs.
Limitations and Future Directions
BRAINSTORMER is an exploratory prototype tested on a limited set of examples. Future work must address whether the knowledge required for a vast library of proverbs remains manageable. Additionally, as the vocabulary expands, more sophisticated search techniques beyond simple forward-chaining may be necessary to maintain tractability. Finally, while the system aims for "instructability," advice must currently be manually encoded into the system's internal representation language.
Summary updated