Introduction
This classic 1986 paper from the AAAI proceedings introduces CHEF, a case-based planner that doesn't just follow recipes—it learns from its own kitchen nightmares. While contemporary AI often relies on massive datasets to learn patterns, Kristian Hammond’s work demonstrates how a system can learn a general rule from a single failure by understanding the "why" behind the mistake.
The Challenge: Anticipation in Planning
In the mid-80s, most automated planners were reactive: they would generate a plan, wait for it to fail due to unforeseen goal interactions, and then attempt to "patch" the error. These systems were "forgetful"—they would often make the same mistake in slightly different contexts because they lacked a mechanism to learn from the failure itself.
The challenge Hammond tackles is anticipation. How can a planner look at a new set of goals (e.g., "make a soufflé with Kirsch liqueur") and realize it’s about to walk into a trap it encountered previously in a different context (e.g., "the strawberry soufflé that fell because of extra liquid")?
The CHEF Methodology
CHEF utilizes Case-Based Planning (CBP), where new plans are created by modifying old ones. Its novel contribution lies in its failure-handling pipeline:
- Simulation & Detection: After building a plan, CHEF runs a simulation. If the results don't match the goals (e.g., the batter isn't risen), it marks the plan as a failure.
- Causal Explanation: The system performs a "back-chaining" analysis using a library of causal rules to find the root cause. It doesn't just see "strawberry," it sees "chopping fruit -> produces liquid -> disrupts liquid/leavening balance -> prevents rising."
- Thematic Organization Packets (TOPs): CHEF maps the explanation to a TOP—a high-level structure that represents a specific type of planning conflict (e.g., a side-effect of one goal disabling a condition for another).
- Generalization: Using its causal rules, CHEF generalizes the failure. It realizes the problem isn't just "strawberries," but any "fruit" or "liquid spice" that might upset the balance.
- Anticipation via "Demons": The system builds "demons" (activation links) in memory. When a new goal set includes features that match a generalized failure, these demons trigger a warning before planning begins.
Key Insights: Learning from a Single Instance
Unlike inductive learning (which requires many examples to find a pattern), CHEF uses explanation-based learning. Because it understands the causal chain of a failure, it can generalize a rule immediately.
- Indexing by "Problems Avoided": CHEF stores its successful "fixed" plans not just by the goals they satisfy (e.g., "strawberry flavor"), but by the problems they avoid (e.g., "excess liquid"). This allows the planner to retrieve a "liquid-safe" plan when it anticipates a moisture problem.
- Surface vs. Structural Similarity: Traditional case-based systems might look for a recipe that shares ingredients. CHEF looks for recipes that share causal structures. It might choose a beef-and-broccoli stir-fry technique to solve a chicken-and-snow-pea problem because both involve the structural conflict of "meat moisture ruining vegetable crispness."
Key Results and Walkthrough
The paper provides a qualitative walkthrough of CHEF’s performance in the domain of Szechwan cooking.
The Strawberry Failure
CHEF tries to make a strawberry soufflé, fails (it falls), explains the failure (too much liquid), and repairs it by adding more egg white.
The Kirsch Success
Later, when asked to make a Kirsch soufflé, CHEF's "demons" notice that Kirsch is a liquid. It anticipates a fallen soufflé and, instead of using a standard vanilla recipe, it retrieves the "fixed" strawberry recipe as a template, successfully avoiding the failure before it ever happens.
The author notes that after learning from about a dozen such complex interactions, the system becomes significantly more robust in its domain.
Limitations and Future Work
- Domain Specificity: The system relies heavily on a pre-defined, accurate causal model of the world (cooking). If the causal rules are wrong or incomplete, the "explanation" will be flawed.
- Scalability: While it works for a "dozen examples," the paper doesn't explore how the "demon" activation system scales to thousands of potential failure types without becoming computationally expensive.
Conclusion
CHEF demonstrates a powerful paradigm shift from "create and debug" to "anticipate and avoid." By performing a causal analysis of its own planning failures, it learns to recognize problematic feature combinations and retrieves specialized plans pre-designed to avoid specific pitfalls.
Summary updated