 |
|
|
Bruce D'Ambrosio
Research Activities
Research Areas
Problem-Solving Architectures for Dynamic Environments, Dynamic, Interleaved,
Construction and Evaluation of Context-Sensitive Decision Bases, Task-Oriented
Approaches to Uncertainty Management for Real-Time Control
Research Description
The fundamental problem facing an autonomous intelligent agent
is to act in the face of limited knowledge and computational resources.
Bayesian or subjective decision theory is arguably the most comprehensive
theory of decision making. Until recently, though, it has had little
application in AI due to computational complexity considerations. Recent
breakthroughs in both representation and inference have dramatically
enlarged the domain of applicability of Bayesian methods in AI. My research
group is active in three areas related to Bayesian decision theory:
(1) representation and inference, (2) inference under time bounds, and
(3) applications, especially in decision support.
The fundamental breakthrough that ignited the current wave of activity
in Bayesian methods in AI was a representation that made explicit, certain
structural aspects which current representations capture poorly or not
at all. Our work on representation and inference is aimed at finding
efficient formalisms for capturing these structural aspects, and inference
algorithms which can exploit them.
Decision theory was developed under the assumption that an agent had
infinite computational resources available. Realizable agents, however,
are severely limited in the amount of information they can store and
the amount of computation they can perform. We are developing methods
by which realizable agents can tradeoff amount of computation against
quality of a resulting decision, as well as methods by which an agent
can make such act/think more. Finally, we believe that Bayesian decision
theory is a rational substructure for many common problems, from machine
diagnosis to design-team decision making. Our group is investigating
a range of problems in collaboration with other groups on and off campus
(e.g., design team decision support with Mechanical Engineering, reforestation
management with Forest Science, and machine monitoring and diagnosis
with Hewlett Packard).
Applications of Research
Current expert systems are capable of serving as useful assistants in
some design and diagnostic tasks. However, they are costly to build,
and limited to simple static situations. Results of this research will
be to extend the range of intelligent assistant systems to include direct
control of engineered systems such as vehicles and manufacturing processes.
Recent Research Collaborations & Projects
- SERC: "Predicting Software Reliability," 1998, $38,000
- NSF: "Real-Time Probabilistic Inference," 1997-2000, $240,000
- SERC: "Predicting Software Reliability," 1997, $15,000
|
 |
