Breakout Sessions: Workshop on Opportunities and Challenges in Uncertainty
Quantification for Complex Interacting Systems
Four breakout sessions are scheduled during the Workshop, two on each
day of the meeting.
The outcome of the breakout sessions will be edited into a report that
will be submitted to NSF and distributed to the participants. A set
of questions are proposed to frame the discussions during the breakout
sessions. Very few and brief presentations by participants (5-10 minutes) are
possible during the sessions.
First Set of Breakout Sessions: Context
The purpose of these breakout sessions is to delineate the societal
and scientific contexts in which several complex systems operate. We
group these systems into three categories, Physics-based models,
Network Models, and Social Sciences. Each of these categories will be
discussed in a separate breakout session. The specific questions to
be addressed within each of these sessions are as follows:
Moderators: Habib Najm (Physics Based Models), Demetri Spanos (Networks), Mark Rintoul (Social Networks)
- How is complexity manifested and/or defined ?
- can nonlinearities be isolated and characterized ?
- can interfaces/linkages be identified (sources of complexity
and uncertainty; junctions where information is exchanged/polluted).
- How is uncertainty manifested/ interpreted ?
- lack of knowledge, lack of data, parametric uncertainty ...
- what are the objectives of uncertainty analysis (what decisions are we trying to make ?)
- What are sources of prior knowledge ?
- accumulated knowledge, behavioral psychology,...
- What are sources of real-world knowledge ?
- physical evidence, opinion polls,...
- how/What data is collected; are experimental resources available/sufficient ?
- How is hybrid data aggregated ?
- multiscale data: patient-specific vs. group/average information.
- multimodel and multiphysics data
- Too much data or too little data ?
- is there a need for a multidisciplinary scientific community to address issues related to Certifiably Rational Approach to Complex Systems ?
Second Set of Breakout Sessions: Challenges
The purpose of this second set of breakout sessions is to identify
challenges in modeling, algorithms and computation associated with
uncertainty quantification, modeling, and large-scale data analysis.
Specifically, the following questions will be addressed.
Breakout session on UQ
Moderator: Gianluca Iaccarino
Scribe: Maarten Arnst
- How is uncertainty tied to complexity ?
- is complexity induced by uncertainty ?
- does complexity overshadow uncertainty ?
- Which uncertainty models are commonly used ?
- probabilistic/fuzzy/evidence/...
- multivariate statistics, polynomial chaos, random matrix ...
- How are decision processes sensitive to uncertainty in uncertainty models (missing data, missing physics, computational resources etc...) ?
- Is data interpretation an issue (credibility of data) ?
- How is observed evidence transformed into model parameters: Inverse analysis ?
- Is high-dimensionality of mathematical analysis or
computational representations an issue ?
- Is there a need for real-time data analysis ?
- Statistical interfaces for (petascale) data-driven applications
- Is there a need for a multidisciplinary scientific
community to address issues related to UQ in Complex Systems ?
Breakout session on Mathematical Modeling
Moderator: Bruce Pitman
Scribe: Sonjoy Das
- Where are models tied to reality ?
- What are the intrinsic features/behaviors that mathematics
attempts to model (i.e. what are the objective functions and what are the constraints) ?
- What mathematical models are used to describe complexity ?
- Are standard assumptions too restrictive ?
- How is complexity characterized ?
- is it intrinsic to the dynamics ?
- does it depend on interaction between systems ?
- does it depend on available information ?
- What is the status of error analysis for the mathematical models used ?
- a-priori/a-posteriori error indicators and error estimators.
- sampling distributions for statistics.
- What are the challenges in stochastic inverse analysis for interacting complex systems ?
- large-scale /real-time optimization challenges
- model reduction
- does an information-rich/driven environment mitigate these challenges ?
- What are the multiscale aspects of complexity and uncertainty that are relevant to the analysis of some relevant systems ?
Breakout session on Large Scale Computational Resources
Moderator: Kevin Long
Scribe: Kristi Potter
What are computational/computer science challenges with complex systems ?
- Is there value for data-driven models in complex systems ?
- Is there value for sensor networks in complex systems ? (for characterization and/or prediction)
- What are the unique features of data management for interacting systems ?
- What are some challenges in large-scale computational statistics ?
- What are some challenges in computational management of hybrid systems (e.g. agent-based models coupled with physics-based models, georeferencing ...)
Third Set of Breakout Sessions:
Breakout session on Mathematical Modeling
Moderator: Susanna Still
- do the interfaces between interacting systems require explicit modeling ? Are they modeled as constraints ? How can they be modeled (PDEs, time series, algebraic constraints ...) Are these interfaces a significant (dominant) source of uncertainty ?
- does a mathematical framework exist for data-driven predictive models (something "like" feedback control ?)
- how about data-driven predictions while quantifying uncertainties ?
- Does a methodology currently exist for end-to-end error analysis (all sorts of error) and management for each system ? does such a methodology seem achievable or are there essential difficulties ?
- what are some critical issues of reduced order models (ROM) for complex systems (capturing some/all emergent behavior?)
- what are some critical issues of ROM for interacting systems (ROM for each subsystem, or ROM for whole system ?) are there some clear mathematical challenges ?
- what are some standards and mathematical challenges in dealing with hybrid systems (continues/discrete, PDEs/ABM/ADE, stochastic/deterministic ...)
- are there outstanding mathematical challenges for developing approximations on graphs (for network applications ?)
- when do heuristic approaches fail and how can they be remedied (more rigor, first principles ...?)
- what set of courses/knowlegde (maths/physics) constitutes foundation for research in complex systems.
Breakout session on Networks
Moderator: Stanley Wasserman
- what are some common multiscale approaches for network modeling ?
- how is experimental evidence assimilated into networks models (calibration)? Sometimes data is available on particular scale(s)
- are there some standard metrics for ascertaining performance of network models ?
- are there outstanding mathematical challenges for developing approximations on graphs (for network applications ?)
- how does network modeling depend on what is being transported on the network (social agents or cars or electrons) ?
- How restrictive/realistic are common models of network models ?
- what set of courses/knowlegde (maths/physics) consistutes foundation for research in network science.
Breakout session on Policy Making
Moderator: Youssef Marzouk
- Give a list of typical policy decisions.
- How relevant is prediction/anticipation to the policy making process ?
- Should politics be modeled as part of the complex system of systems ?
- Can the case be made for UQ standards in communicating information to policy makers ? what are current standards ?
- How complex/sophisticated can a decision maker be ?
- Are the research funding avenues well aligned with the research needs? Do the agencies need to create new "windows" or do the existing ones suffice?
Fourth Set of Breakout Sessions: Opportunities
Moderator : Roger Ghanem
- What are the top 3 research questions a) in UQ for complex systems? b) UQ for characterizing interactions?
- What are the primary needs for tools and infrastructure (software, analysis procedures ...)
- Are the research funding avenues well aligned with the research needs? Do the agencies need to create new "windows" or do the existing ones suffice?