TL;DR
This paper introduces a new perspective on artificial intelligence (AI) systems that deal with uncertainty, shifting the focus from merely assessing belief to actively managing uncertainty through planning. The authors present MUM and its successor MU, architectures designed to generate sequences of actions (like questions, tests, or treatments) that not only achieve domain goals but also strategically minimize uncertainty and its consequences.
What Problem Does This Paper Address?
Traditional AI approaches to uncertainty often focus on calculating the probability or belief in different hypotheses given evidence. However, this doesn't tell a system what to do when it's uncertain. Imagine a doctor faced with a patient exhibiting symptoms that could indicate one of two diseases. While they might be able to assign probabilities to each disease, the crucial question is: what action should they take next? This paper addresses this gap by proposing that managing uncertainty is fundamentally a planning problem. The challenge is to design systems that can make decisions and take actions even when faced with incomplete information, where these actions are chosen not just for their direct impact on the domain goal (e.g., treating a disease) but also for their effect on reducing uncertainty itself.
Methods
The core of the paper's contribution lies in its architectural proposals for managing uncertainty.
- MUM (Managing Uncertainty in Medicine): This system was designed to generate diagnostic "workups" – sequences of questions, tests, and treatments. MUM uses a large inference network where nodes represent diseases, clusters of evidence, and raw data. It reasons about the utility of actions by considering their potential to confirm or detract from hypotheses, their costs, and their impact on uncertainty. A key aspect is the use of discrete belief levels (e.g., confirmed, strongly-supported, unknown) to simplify local combining functions.
- MU Architecture: Building on MUM's experience, MU is presented as a more general and flexible architecture. It further generalizes the inference network and allows for control parameters beyond just belief levels to be propagated through local combining functions. MU provides a language for experts to define the state of a problem and query aspects of that state. This allows for declarative specification of control strategies, making it easier to build and modify systems that manage uncertainty. MU supports five classes of queries about the problem state, enabling sophisticated reasoning about actions and their consequences.
Datasets
Not discussed in the paper. The paper focuses on architectural design and reasoning principles, using medical diagnosis and plant pathology as illustrative domains.
Key Insights and Ideas
- Uncertainty as a Constraint and Facilitator: Counterintuitively, uncertainty can actually help in problem-solving by constraining the set of possible actions. For example, if two disease hypotheses suggest overlapping treatments, uncertainty about which disease is present might force the system to choose a treatment that covers both, simplifying the decision.
- Actions Require and Produce Evidence: In this framework, actions are not just about achieving domain goals; they are also mechanisms for gathering more information. Every action, whether a question, a test, or a treatment, can provide evidence that updates the system's beliefs.
- Certainty is Not Always a Prerequisite for Action: The paper argues that systems can and should act even when certainty is not achieved. The level of evidence required for an action depends on its cost and potential consequences. A less costly action might be warranted with weaker evidence than a more expensive or risky one.
- Local Combining Functions for Belief: The authors advocate for local combining functions for belief updates, arguing they are more modular and easier to modify than global approaches. This allows for incremental system development and maintenance.
- Managing Uncertainty is Integral, Not an Add-on: The paper emphasizes that managing uncertainty is not a separate module but an intrinsic part of the problem-solving process, influencing action selection at every step.
Key Results
The paper doesn't present quantitative results from empirical evaluations of MUM or MU in the traditional sense. Instead, it focuses on the conceptual framework and architectural design. The key "results" are the proposed architectures (MUM and MU) and the demonstration of how they can be used to generate expert-like workups in medicine. The paper highlights the successful reimplementation and modification of MUM's control structure using the MU architecture, suggesting its flexibility and effectiveness.
Limitations and Future Work
- Knowledge Engineering Effort: While MU aims to simplify control specification, the knowledge engineer still needs to define control strategies. The authors acknowledge that developing a catalog of control strategies for various uncertain situations will be a future endeavor.
- Generality Testing: The paper states that MU's generality will be tested by implementing several systems in it over the next few months.
- Focus on Tools: The current focus is on building the tools (MU architecture) to acquire and manage uncertainty, rather than on developing a comprehensive library of control strategies.
The authors envision future work involving the development of a catalog of control strategies appropriate for different types of uncertain situations, building upon the foundation laid by the MU architecture.
Summary updated