Abstract
Traditional AI planning systems often struggle with multi-constraint problems, typically generating individual plans for separate goals and subsequently resolving conflicts through meta-level reasoning. This paper introduces WOK (Well Organized Knowledge), a case-based planner that shifts the paradigm from "building from scratch" to "adapting from memory." By organizing knowledge around the interactions between goals rather than the goals themselves, WOK demonstrates a more human-like approach to solving complex problems within the domain of Szechuan cooking. This paper explores how indexing plans by the conflicts they resolve allows for more efficient and robust plan generation.
Introduction
In the early 1980s, AI planning was dominated by systems designed to satisfy single goals or linear sequences of goals. When faced with multiple, potentially conflicting constraints—such as a dish that must be both spicy and sweet—traditional models like NOAH or Wilensky’s planners would generate independent plans for each constraint and then utilize "critics" to fix the resulting clashes.
Kristian Hammond argues that this "plan-then-fix" approach is computationally inefficient and fails to capture the essence of human expertise. Humans rarely reinvent solutions; instead, we recall past situations where we successfully balanced similar competing needs and adapt those solutions to the current context. WOK was developed to address the challenge of indexing and retrieving plans based on the relationships between goals.
The WOK Methodology
WOK is a case-based reasoning (CBR) system built upon Roger Schank’s theory of Thematic Organization Packets (TOPs). The system operates through a four-stage process designed to prioritize goal interactions over simple ingredient matching.
1. Goal Sorting and Interaction Identification
When presented with a set of user constraints (e.g., specific ingredients, desired tastes, or textures), WOK first ranks these goals by their "importance." Importance is determined by how frequently a goal participates in significant interactions within the system's database. Once the top goals are identified, WOK characterizes the abstract interaction between them, such as CONTRASTING-TASTE: EQUAL-FOOTING.
2. Interaction-Centric Retrieval
Unlike a standard database search that might look for a specific ingredient like "chicken," WOK searches its memory for a recipe that successfully manages the identified interaction. For example, if a user requests a dish with chicken and oranges, WOK might retrieve a recipe for "Pork with Hoisin." While the ingredients differ, the underlying plan is relevant because it successfully balances savory and sweet tastes of equal intensity.
3. Plan Modification
Once a skeletal plan is retrieved, WOK maps the new goals onto the old structure. This involves simple replacements (pork to chicken) and functional checks to ensure the new ingredients can fulfill the roles of the originals (e.g., can the new ingredient be stir-fried in the same manner?).
4. Failure Avoidance
Finally, WOK subjects the new plan to a "critic" phase that draws on a library of past failures. The system checks for known "bad" interactions—such as a previous attempt where oranges and scallions resulted in an unpalatable flavor. If a potential failure is detected, WOK applies general strategies, such as introducing a "buffer" ingredient or a different contrast, to preemptively fix the plan.
Case Study: Chicken and Mandarin Oranges
The efficacy of WOK is illustrated through its generation of a recipe for Chicken and Mandarin Oranges.
- Retrieval: WOK identifies the "Savory vs. Sweet" interaction and retrieves the "Pork Shreds with Hoisin" case.
- Adaptation: It substitutes chicken for pork and mandarin oranges for hoisin sauce.
- Correction: During the failure avoidance phase, the system recalls a failure from an "Orange and Olive Salad" case, noting that scallions (present in the original pork recipe) clash with oranges.
- Refinement: Consulting its general strategies for
CONTRASTING-TASTES, WOK determines that "Red Pepper" provides a superior contrast to chicken and oranges. The final output is a successful, novel recipe: a variation of Chicken with Peanuts and Red Peppers.
Key Insights
Interaction-Centric Indexing
The core contribution of WOK is the idea that the most powerful way to store a plan is by the conflict it resolves. Knowing a recipe handles "salty vs. licorice" is more computationally useful for general planning than simply knowing it contains "soy sauce."
Planning as Memory Reminding
WOK treats planning as a memory task. By being "reminded" of a similar past success, the system gains a skeletal structure that bypasses the need for exhaustive search or complex simulations from first principles.
Failure-Driven Refinement
By explicitly storing "negative cases," WOK can anticipate and bypass known pitfalls. This allows the planner to avoid time-consuming simulations by using memory to predict where a plan might fail before it is even finalized.
Conclusion and Future Work
While WOK's implementation is specific to the rules of Szechuan cooking and relies on manually engineered knowledge, its underlying principles are broadly applicable. The system demonstrates that by focusing on goal interactions and utilizing episodic memory, AI planners can achieve more efficient and human-like problem-solving. Future directions for this research include the autonomous learning of taste categories and scaling the library of interactions to handle more complex, multi-domain planning tasks.
Summary updated