This 1987 paper from the AAAI proceedings introduces MU, a development environment designed to help researchers build systems that can "feel their way" through a problem. Unlike traditional AI planners of the era that assumed perfect knowledge, MU focuses on prospective reasoning: the art of deciding what to do next when the world is uncertain, information is expensive, and every action has trade-offs.
The Challenge of Uncertainty in Planning
In the mid-1980s, most AI planning followed a "dead reckoning" approach, where planners assumed the initial state was fully known and that actions had predictable, deterministic effects. However, in fields like medical diagnosis, a physician doesn't have a complete map of a patient's internal state; they perform tests to reveal it. These tests have costs—money, time, patient discomfort—and varying degrees of "diagnosticity."
MU was built to solve the rigidity of previous systems (like MUM) where diagnostic strategies were buried in complex code. By making strategic "control features" explicit and easy to manipulate, MU allows developers to define how a system should prioritize a "dangerous" disease over a "cheap" test without days of debugging.
The MU Architecture
MU provides a task-level architecture that separates domain knowledge from the planning strategy. Its core component is a frame-based Inference Network where raw data feeds into "clusters" (groups of symptoms), which then provide evidence for "hypotheses" (diseases).
Within this network, MU manages four classes of features:
- Static: Fixed values like the monetary cost of a blood test.
- Datum: Values acquired at run-time by asking the user.
- Dynamic: Values computed via "combining functions" (e.g., calculating the current "level of support" for a diagnosis based on incoming test results).
- Focus: Sets of nodes that meet specific criteria, such as "all triggered hypotheses that aren't yet confirmed."
The Planner-MU Interface
MU does not include a pre-built planner. Instead, it acts as a specialized environment that provides a query language for a developer's planner. This interface allows the planner to ask questions about the current state, explore hypothetical outcomes, and request data from the user.
Crucially, the "combining functions" can be used prospectively. A planner can ask: "If I were to get a negative result on an EKG, how would that change my belief in Angina?" This allows the system to evaluate the utility of evidence before committing to an action.
Key Insights and Results
- "Feeling Your Way" vs. Dead Reckoning: The authors argue that in uncertain domains, plans must be short. You take an action, see what happens, and re-evaluate.
- Explicit Control Features: Instead of hiding decision logic in code, MU forces the developer to define features like criticality, dangerousness, and evidence utility. This makes the system's "thinking" transparent.
- Development Speed: The most significant qualitative result was the reduction in development time. The authors successfully re-implemented the MUM system (a program for chest and abdominal pain) within MU. What previously took days of Lisp programming to change a diagnostic strategy could now be done by defining new control features and combining functions within the MU interface.
Limitations and Future Work
While MU simplifies the representation of strategies, acquiring those strategies from experts remains a bottleneck. Experts often struggle to provide precise "combining functions" for complex data. The authors suggest future work in automated strategy learning—inducing strategies from case studies—and refining the interface to include sensitivity analysis, allowing a planner to identify which missing data the current belief is most sensitive to.
Summary updated