Abstract
Automated planning in complex, knowledge-rich domains faces a significant computational hurdle: the combinatorial explosion of goal conflict detection. The Knowledge Intensive Planner (KIP) addresses this by introducing the concept of "concerns." By focusing only on likely sources of plan failure rather than exhaustive cross-checking of every effect against every goal, KIP allows AI consultants to provide safe, context-aware advice. This paper details the KIP architecture within the UNIX Consultant (UC) system, demonstrating how structured concern management enables efficient and human-like planning.
The Challenge of Goal Conflict Detection
In automated planning, Goal Conflict Detection (GCD) is often a nightmare of scale. Consider a UNIX consultant advising a user on moving a file. If the system selects a plan (e.g., the mv command), it must theoretically ensure the plan is "safe" by checking every consequence against every possible user goal.
In a typical knowledge-rich environment, a single command might have 5 effects, while a user might have 50 background goals (e.g., "don't delete important data," "don't exceed disk quotas"). A brute-force approach requires 250 checks for a single simple command. Humans avoid this overhead by ignoring irrelevant information and focusing only on "concerns"—specific ways a plan is likely to fail in a given context. KIP aims to replicate this common-sense focus in AI.
The KIP Concern Algorithm
KIP, a core component of the UNIX Consultant (UC) system, manages planning through a structured, three-way relationship between a plan, a conflicting effect, and a threatened goal. The system follows a multi-step algorithm to process these concerns efficiently.
1. Retrieval and Filtering
Instead of searching all possible goals, KIP retrieves "stored concerns" indexed to the specific plan under consideration. It then applies a taxonomy filter to distinguish between different types of conflicts:
- Default vs. Violated-Default: For instance, printing a standard file is routine, but printing a "very long file" triggers a violated-default concern regarding shared resource usage.
- Intended vs. Unintended Effects: KIP avoids generating "artificial conflicts." If a user explicitly asks to delete a file, KIP will not warn them that the file will be deleted. However, if a plan to "free up space" incidentally deletes a file the user intended to keep, a concern is raised.
2. Evaluation and Treatment
KIP assigns a "degree of concern" (on a scale of 1-10) based on the likelihood of the effect and the importance of the threatened goal. Depending on this score, the system chooses one of three paths:
- Disregard: Low-priority concerns are ignored to maintain efficiency.
- Overlook: Moderate concerns are noted; the plan proceeds, but the user is warned in the final output.
- Elevate: High-priority concerns are treated as plan failures, triggering a search for a new, safer plan.
Key Insights and Knowledge Representation
KIP’s efficacy relies on several sophisticated distinctions in its knowledge representation:
- Interests vs. Goals: An "interest" is a general state (e.g., "I want to keep my files"), while a "goal" is a concrete instance (e.g., "I want to keep
report.txt"). KIP only instantiates a concrete goal when an interest is specifically threatened, saving memory and processing power. - Contextual Relevance: The system recognizes that the same command requires different levels of caution depending on user intent.
- Inheritance of Concerns: Because plans are organized hierarchically in the KODIAK knowledge base, specific commands (like
lprfor a large file) can inherit concerns from broader categories (like "resource-intensive-plans"), allowing for robust generalization.
Results and Qualitative Performance
KIP was evaluated using the KODIAK Knowledge Base, which contains a hierarchy of UNIX commands and user interests. Expert knowledge was manually encoded to populate the system's initial set of concerns. Qualitative traces from the UNIX Consultant system demonstrate KIP's success in several scenarios:
- Password Synchronization: KIP detects that changing a password on one machine might conflict with a user's interest in maintaining identical passwords across a network.
- Resource Management: The system correctly identifies when a massive print job violates social defaults of system consideration.
- Safety Warnings: When a user attempts to move a file to an existing destination, KIP identifies the potential deletion of the destination file and appends a warning: "However, if file1 exists it will be deleted."
Limitations and Future Directions
The primary limitation of the current KIP implementation is the reliance on manual encoding; all concerns must be authored by human experts. Future work suggests a learning component to automatically detect the frequency of plan failures from user interactions. Additionally, while KIP is more efficient than brute force, further research is needed to quantify how the system scales as the library of concerns grows into the thousands.
Summary updated