Introduction
In the mid-1980s, the dominant paradigm in artificial intelligence planning relied on "first principles." Systems like STRIPS constructed plans by chain-linking rules and operators from scratch for every new goal. While logically sound, this approach was computationally expensive and fundamentally "forgetful." These planners lacked the ability to learn from experience, often repeating the same tactical errors and requiring manual debugging for every new instance of a problem.
Kristian Hammond’s CHEF system proposed a radical alternative: Case-Based Planning (CBP). Hammond argued that human planning is episodic rather than purely algorithmic. When we cook, we do not re-derive the laws of thermodynamics; we remember what worked—and what failed—in the past. CHEF was designed to demonstrate how a machine could use its own history to anticipate and avoid failures before they occur.
The Case-Based Planning Cycle
CHEF operates through a six-stage cycle that integrates planning, execution (via simulation), and learning. Unlike traditional planners that focus solely on goal satisfaction, CHEF’s architecture is built around the management of expectations.
- Problem Anticipation: Before a plan is even retrieved, CHEF analyzes the user's goals to identify potential conflicts. It looks for combinations of features that have historically caused failures.
- Plan Retrieval: The system searches its memory for a "best match" recipe. Crucially, it does not just look for ingredient similarity; it prioritizes plans that are indexed as having successfully avoided the problems anticipated in the first step.
- Plan Modification: Once a baseline plan is found (e.g., a beef and broccoli recipe), CHEF performs substitutions to meet the current goals (e.g., replacing beef with chicken).
- Plan Repair: If the modified plan fails during simulation, CHEF uses a causal model of the domain to explain the failure.
- Credit Assignment: After explaining a failure, CHEF identifies the specific combination of features responsible. It then creates a "demon"—a background monitor—to predict this specific failure in future scenarios.
- Plan Storage: Successful or repaired plans are stored in memory, indexed by both the goals they satisfy and the specific failures they are designed to avoid.
Learning Through Failure
A central insight of the CHEF model is that failure is a powerful learning signal. In this framework, a failure is viewed as an "expectation failure," indicating that the planner’s internal model of the world is incomplete.
To generalize these lessons, CHEF utilizes Thematic Organization Packets (TOPs). TOPs are high-level structures that describe abstract planning interactions. For example, a TOP might describe a "Side-effect: Disabled-condition: Concurrent" interaction, such as when the moisture released by one ingredient (meat) prevents another ingredient (vegetables) from staying crisp. By abstracting the failure, CHEF can apply the lesson to entirely different ingredients that share the same causal properties.
Domain: Szechwan Cooking
CHEF was implemented in the domain of Szechwan cooking. This domain provided a rich set of interacting constraints and causal relationships. The system’s knowledge base included the "physics" of the kitchen—understanding how heat and moisture interact—while its case library grew as it experimented with new recipes.
Key Results and Evolution
The paper illustrates CHEF’s logic through a specific evolutionary sequence:
- Initial Failure: CHEF attempts to create a beef and broccoli stir-fry. It stir-fries both together, but the liquid released by the beef makes the broccoli soggy.
- Explanation and Repair: The system explains the failure (excess liquid + crisp vegetable) and repairs the plan by stir-frying the components separately and combining them at the end.
- Indexing by Avoidance: CHEF stores this new plan, indexing it as a way to avoid "sogginess" when dealing with meat and crisp vegetables.
- Anticipation: When later asked to prepare chicken and snow peas, CHEF anticipates the sogginess problem. Even if it has a "chicken and green beans" recipe that is a closer surface match, it chooses the "beef and broccoli" plan as its baseline because that plan is specifically marked as a solution to the anticipated failure.
Limitations and Conclusion
While CHEF represented a major leap in AI learning, it was not without limitations. The system relied on a hand-coded causal model of its domain, making it difficult to scale to "open-world" environments without significant knowledge engineering. Furthermore, its "execution" phase took place in a controlled simulator; real-world planning involves noise and unpredictable variables that complicate plan repair.
Despite these constraints, CHEF proved that indexing experiences by the problems they solve—rather than just the goals they meet—allows a planner to become more efficient and robust over time. It shifted the focus of AI planning from pure logic to the effective use of memory.
Summary updated