Multiple Model Approaches in Decision Making
Motivation
As the society advances into the future, natural and engineered systems are being pushed to operate under increasing degrees of uncertainty. The large sources and types of uncertainties, along with presence of conflicting and/or subjective criteria that arise from different stakeholders, has made the task of planning and decision-making even more arduous in many sectors.
Take, for example, the socio-technical problem of integrated management of river basins and watersheds. Whether it is management of floods, droughts, water quality, or of ecological systems, the management of watersheds requires a rigorous understanding of not only the physical and biological laws that govern how water is stored, regulated, and routed in the natural and built environments of the watershed, but also cognitive and social processes related to watershed stakeholders (i.e., urban and rural landowners, small and large municipalities, managers, watershed councils, and agency personnel, etc.) who influence what type of infrastructure solutions are implemented and where. Therefore, design of management solutions for watersheds requires stakeholders such as farmers, environmentalists, and government officials, to be cognizant of the presence of diverse sources of uncertainties arising from coexisting physical and human systems, and generate decisions that include community criteria beyond the usual physical and socioeconomic constraints.
Traditional Systems Analysis approaches of developing a model and using it to make decisions for socio-technological systems are proving inadequate since a single model is unable to cope with large uncertainties. In various disparate areas of Systems Theory and Decision Sciences, multiple model based methods (referred to variously as “multi-model methods”, “ensemble methods”, “multi-classifier systems”, “mixture of experts”, “multi-agent systems”, etc.) are increasingly being used to deal with large uncertainties in physical and human dimensions. However, an integrated framework for analyzing and synthesizing such diverse multiple model based approaches in diverse disciplines is still lacking.
Purpose
The objective of the workshop is to bring together leading researchers and practitioners from multiple disciplines to discuss the state-of-the-art of both the theory and practice of multi-model based approaches for physical and human systems, define new research directions to advance this field, and formulate an integrated, interdisciplinary view of multiple models based methods for decision making under multiple conditions of uncertainties.
Products
All the attendees will be invited to be co-authors on a survey report/white paper on the state of the art and research needs in multi-model methods, planned to be submitted to a large-impact venue (such as the “Science” journal) for possible publication. Other potential concrete results may include development of new research collaborations and potential future external/federal research grant proposals.
Workshop Details
The overarching goal of this proposed multi-disciplinary workshop will be to discuss and share ideas about novel, mathematical and computational frameworks for characterizing and coping with large uncertainty via multi-model methods. Theoretically, multiple models can be developed for both the physical/engineering systems as well as for the human decision makers and stakeholder to present uncertainty in the coupled physical-human design environment. However, practical challenges remain in using such approaches in applications involving coupled and complex human-physical systems.
The proposed workshop will be virtual to facilitate remote participation of researchers and practitioners from diverse disciplines, including civil and environmental engineering, computer science, control engineering, mathematical systems theory, and decision sciences. The workshop will consist of the following four sessions involving presentations from invited speakers and group discussions among attendees, distributed over two days.
Go to TopSchedule
Tuesday May 30, 2023
| Time (Eastern Time) | Event | Speakers |
|---|---|---|
| 12:00 PM - 12:20 PM | Welcome | Dr. Meghna Babbar-Sebens, Associate Professor of Water Resources Engineering |
| Session 1 12:20 PM - 1:00 PM |
Panel 1 |
Theoretical Foundations of Multiple Model Methods Dr. Kumpati S. Narendra and Dr. Cheng Xiang Moderator: Dr. Snehasis Mukhopadhyay |
| 1:00 PM - 1:45 PM | Breakout Session 1 | |
| 1:45 PM - 2:15 PM | Break | |
| Session 2 2:15 PM - 2:55 PM |
Panel 2 |
Applications of Multiple Models for Uncertainty in Physical (Natural and/or Engineered) Systems Dr. David Rupp (Prism Climate Group) and Dr. Raghavan Srinivasan Moderator: Dr. Meghna Babbar-Sebens |
| 2:55 PM - 3:40 PM | Breakout Session 2 | |
| 3:40 PM - 4:00 PM | Summary of Day 1 | |
Wednesday May 31, 2023
| Time (Eastern Time) | Event | Speakers |
|---|---|---|
| 12:00 PM - 12:20 PM | Welcome | Dr. Meghna Babbar-Sebens, Associate Professor of Water Resources Engineering |
| Session 3 12:20 PM - 1:00 PM |
Panel 3 |
Applications of Multiple Models for Uncertainty in Human Systems Dr. John O'Doherty and Dr. Cleotilde (Coty) Gonzalez Moderator: Dr. Kristen Macuga |
| 1:00 PM - 1:45 PM | Breakout Session 3 | |
| 1:45 PM - 2:15 PM | Break | |
| Session 4 2:15 PM - 2:55 PM |
Panel 4 |
Multiple Models - Role of Artificial Intelligence and Cyberinfrastructure Dr. Arjan Durresi and Dr. Sudhakar Pamidighantam Moderator: Dr. Suresh Marru |
| 2:55 PM - 3:40 PM | Breakout Session 4 | |
| 3:40 PM - 4:00 PM | Summary of Day 2 |
Go to Top
Session Abstracts
Day 1: Tuesday May 30, 2023
Session 1
Panel 1 Title: Theoretical Foundations of Multiple Model Methods
Speaker 1: Dr. Kumpati S. Narendra
Title: Adaptive Control Using Multiple Models
Abstract
Over the past sixty years, adaptive methods have been well established for the identification and control of linear time-invariant systems with unknown parameters using a single model. However, problems encountered in recent years in biology, medicine, neuroscience, vision and economics are nonlinear, with rapidly time-varying parameters and stochastic disturbances. Classical adaptive control using a single model has been found to be inadequate in these cases and new methods that are significantly faster are needed. The methods based on multiple models developed during the past twenty years attempt to achieve this objective. The lecture will discuss the evolution of the different improvements that have been proposed, including the use of multiple fixed identification models, switching and tuning, the use of multiple identifiers and controllers for improved transient response, and finally second level adaptation.
Biography
Dr. Kumpati S. Narendra is currently the Harold W. Cheel Professor of Electrical Engineering and the Director of the Center for Systems Science at Yale University.
He received his Ph.D. degree from Harvard University and was an Assistant Professor there until 1965, when he came to Yale. He was made Professor in 1968. He was the Chairman of the Electrical Engineering Department from 1984 to 1987 and Director of the Neurongineering and Neuroscience Center during 1995-96. He received an Honorary Doctorate from his alma mater, the University of Madras (now Anna University) in 1995, and the University of Ireland, Maynooth in 2007.
Professor Narendra is the author of over 250 papers and 3 books and the editor of 4 books in the fields of Adaptive Control, Learning Theory, Stability Theory and Neural Networks. He has delivered over 100 lectures in 45 countries and has mentored 49 PhD students and 38 postdoctoral and visiting fellows in the past 40 years. During the same period he was a consultant for over 15 industrial research laboratories, including General Motors, AT&T, Sikorsky Aircraft and Sandia National Labratories.
Professor Narendra was the recipient of the Franklin V. Taylor Memorial Award of the Systems, Man, and Cybernetics Society of the IEEE (1972), the George Axelby Best Paper Award of the Institute of Electrical and Electronics Engineering (IEEE) Control Systems Society (1987), the John R. Ragazzini Education Award (1990), the IEEE Neural Networks Society Best paper Award (1991), the Bode Prize (1995), the Walton Fellowship in Ireland (2007) and the Neural Networks Pioneer Award (2008).
In 2003, he received the Richard E. Bellman Control Heritage Award, the highest award of the American Automatic Control Council (AACC) for “Pioneering Contributions to Stability Theory, and Adaptive and Learning Theory”.
Speaker 2: Dr. C. Xiang
Title: Data-driven Identification and Control of Nonlinear Systems using Multiple Models and Neural Networks
Abstract
The multiple model framework provides an effective tool for identification and control of nonlinear systems. Among various multiple model structures, the piecewise affine (PWA) models have emerged as the most popular approach due to its mathematical simplicity. However, there are two well-known issues for the PWA approach: the curse of dimensionality and the computational complexity. To overcome these two challenging problems, we have proposed a novel multiple model approach. Different from PWA models in which all dimensions of the regressor space are partitioned, the key idea is to only divide the space of the present control input u(k) at time k (the instant of interest in the controller design) into several intervals and identify a local model that is linear in u(k) within each interval. With the help of universal approximators such as neural networks, the nonlinear terms in the local models can be easily approximated. With the proposed multiple model architecture, a computationally efficient switching control algorithm is derived to control the nonlinear systems based on the weighted one-step-ahead predictive control method and constrained optimization techniques. The upper bound for the tracking error using this switching control strategy can be established under certain assumptions. Both simulation studies and experimental results demonstrate the effectiveness of the proposed multiple model architecture and switching control algorithm.
Biography
Dr. C. Xiang received the B.S. degree in mechanical engineering from Fudan University, China in 1991 and Ph.D. degree in electrical engineering from Yale University in 2000. From 2000 to 2001 he was a financial engineer at Fannie Mae, Washington D.C.. He has been with the National University of Singapore since 2001. At present, he is the Area Director of Control, Intelligent Systems and Robotics at the Department of Electrical and Computer Engineering, the National University of Singapore. His research interests include intelligent control, artificial intelligence and systems biology.
Day 1: Tuesday May 30, 2023
Session 2
Panel 2 Title: Applications of Multiple Models for Uncertainty in Physical (Natural and/or Engineered) Systems
Go to TopSpeaker 1: Dr. David Rupp
Title: Applications of multiple-model frameworks for exploring uncertainty in climate change impacts on water resources
Abstract
This presentation summarizes some work towards using multiple-model frameworks to characterize uncertainty when projecting the impacts of climate change on water resources and dependent sectors (e.g., hydropower). Strengths and remaining deficiencies of the employed methods are discussed.
Biography
Dr. David Rupp is Associate Professor, Senior Research, College of Earth, Ocean, and Atmospheric Sciences, Oregon State University. He studies climate change and its impact on natural and built environments and has investigated the effect of anthropogenic emissions on a range of issues, including extreme precipitation, flooding, drought, heatwaves, wildfire, forest vulnerability, power generation, and water supply reliability. He often uses regional climate models to examine how the interaction of the atmospheric circulation and the land surface determines spatial variability in climate change and its impacts, with a focus on the western United States. For over a decade, David has contributed to the Oregon Climate Research Institute’s mission of providing information and technical assistance on climate change impacts to the public and regional, state, and local government.
Speaker 2: Dr. Raghavan Srinivasan
Title: Sources of Uncertainties in Multi-modeling and its Impacts on Results.
Abstract
Watershed Models suffer from large model uncertainties. These can be divided into conceptual model uncertainty, input uncertainty, and parameter uncertainty. The conceptual model uncertainty (or structural uncertainty) could be of the following situations: a) Model uncertainties due to simplifications in the conceptual model, b) Model uncertainties due to processes occurring in the watershed but not included in the model, c) Model uncertainties due to processes that are included in the model, but their occurrences in the watershed are unknown to the modeler, and d) Model uncertainties due to processes unknown to the modeler and not included in the model either! Input uncertainty results from errors in input data such as rainfall and, more importantly, the extension of point data to large areas in distributed models. Parameter uncertainty is usually caused due to inherent non‐uniqueness of parameters in inverse modeling. Parameters represent processes. The fact that processes can compensate for each other gives rise to many parameters that produce the same output signal. In addition, all these uncertainties combined with observation data used for calibration and validation have their uncertainties. No studies comprehensively studied the impact of propagation of all the uncertainties resulting from input/model/parameter/observed data. The presentation will highlight the challenges we face when modeling watersheds with all sources of uncertainties.
Biography
Key role: Director of the Spatial Sciences Laboratory at TAMU and Resident Director of the Blackland Research and Extension Center. University Distinguished Professor in Department of Ecology and Conservation Biology and Department of Biological and Agricultural Engineering.
R. Srinivasan is a professor at Texas A&M University and director of the Spatial Sciences Laboratory at Texas A&M. He has become known and respected throughout the world for his developmental work with spatial sciences and computer-based modeling, especially the Soil and Water Assessment Tool or SWAT model. His research and its applications have contributed to long-lasting changes in natural resource assessments and development of management system options, currently being used in more than 90 countries. Over the past nine years, he has conducted more than 60 international workshops for students and professionals in more than 20 countries. He was the Principal Investigator or Co-PI on research grants from USAID, EPA, NOAA, DOE, USDA, NASA, NSF. He is a member of the EPA’s advisory group for national water quality assessments and was a member of the TNRCC’s science and technical advisory committee on the proper procedures to adopt TMDL for the state of Texas from 1998-1999. He was appointed as a member of the Scientific Advisory Panel in the FIFRA Scientific, and he served as an associate Editor for Sustainability, Special issue from 2016 -2018.
Dr. Srinivasan is widely recognized for innovatively revolutionizing the user interface, data, and graphical support for the existing hydrological model (Soil and Water Analysis Tool, SWAT) to facilitate use by scientists and decision makers. The resultant innovative tool, called the Hydrologic and Water Quality System (HAWQS), combines scientific elements drawn from multiple domains, including geographic information systems, computer science, database development, data management, and graphical user interfaces. Due to his efforts, SWAT/HAWQS is widely acknowledged as the most comprehensive hydrologic modeling platform used globally to support policy analysis worldwide. Dr. Srinivasan has led and contributed to numerous studies evaluating the hydrological and water quality effects of agricultural production practices, watershed management, climate change, and land-use change along with their implications for agricultural productivity, hydrology, water quality, environmental sustainability, farm family nutrition, and farm economics. Many of these studies include the study of erosion and sediment at the river basin scale, including soil erosion in watersheds, sediment extraction and storage, sediment transport, and riverbank erosion along with mitigation and management of sediment in reservoirs. Dr. Srinivasan's major scientific contribution is the development and advancement of decision support tools that integrate computer models for watershed dynamics, GIS visualization, and big data into a platform that has allowed substantially more scientifically based hydrological analysis while improving interpretation of findings, and providing means for improved communication with managers, regulators, and policymakers. These efforts have transformed the way that watershed scientists and managers address environmental issues, perform assessments, and communicate options.
Day 2: Wednesday May 31, 2023
Session 3
Panel 3 Title: Applications of Multiple Models for Uncertainty in Human Systems
Speaker 1 : Dr. John O'Doherty
Title : Applications of Multiple Models for Uncertainty in Human Systems
Abstract
It has long been suggested that human behavior can be understood as reflecting the contributions of multiple systems that cooperate or compete for the control of behavior. Here we suggest that the brain can be thought of as a “Mixture of Experts” in which multiple different expert systems propose strategies for action. I will consider how the brain determines which system should control behavior at any one moment in time. It will be argued that this is accomplished by keeping track of the reliability of the predictions within each system, and by allocating control over behavior in a manner that is proportional to the relative reliability of those predictions. I will present behavioral and neural evidence for the existence of a reliability-based control mechanism operating over multiple experts in humans. These include model-based and model-free reinforcement-learning strategies that learn to select actions on the basis of direct experience, experts that learn to select actions through observing the behavior of other agents, as well as a system that reflexively takes actions based solely on visual affordances. Results from the study of different expert systems in both experiential and social-learning domains hints at the possibility that this reliability-based control mechanism is domain general, exerting control over many different expert systems simultaneously in order to yield sophisticated and adaptive behavior.
Biography
John P. O'Doherty is the Fletcher Jones Professor of Decision Neuroscience at the California Institute of Technology. He studies the neural basis of reward-related learning and decision making, complex computational problems that have come to be solved by the brain over the course of evolution. He's interested in how the human brain can learn from experience in order to make decisions that maximize future rewards and minimize future costs. To address this question, he utilizes computational models in combination with behavioral, neuroimaging and neurophysiological data collected from human participants. O'Doherty completed a D.Phil at the University of Oxford (2001), and was a postdoctoral scholar at the Wellcome Dept of Imaging Neuroscience at University College London until 2004, after which he joined the California Institute of Technology. He is a fellow of the Association for Psychological Science.
Speaker 2: Dr. Cleotilde (Coty) Gonzalez
Title: Instance-Based Learning Theory of Decisions from Experience in Dynamic Environments
Abstract
Humans make decisions is environments that change over time and under increasing degrees of uncertainty. The field of Dynamic Decision Making (DDM) studies how humans make decisions in such situations and how they learn from past decisions to adapt and improve their choices over time. The most well-known theory of DDM is Instance-Based Learning Theory (IBLT) (Gonzalez, Lerch & Lebiere, 2003). IBLT emerged from a set of behavioral phenomena established in experiments that used Microworlds (complex interactive decision games), and from the efforts to implement computational algorithms that would replicate the human decision process involved. The theoretical process and mechanisms proposed in IBLT have been used in the computational implementation of a large number of models in multiple fields. The power of the predictions of IBL models compared to human decisions have been demonstrated in a large diversity of tasks including control of carbon dioxide in the atmosphere, supply chain inventory management, search and rescue and navigation in non-stochastic situations, and dynamic resource allocation of resources in cyber defense. The talk presents the theoretical principles of IBLT, and illustrates how IBL models have been used to achieve a level of dynamic and adaptive autonomy in cyber defense. Dr. Gonzalez will provide a high-level overview of the advances we have achieved by using a cognitive approach for modeling the attacker's and end-user's decisions; exploring various deception techniques; and performing empirical demonstrations of these deception techniques in tasks of increasing complexity and realism. The talk will conclude with a discussion of the future potential of IBL models in Human-Automation Teams more generally.
Biography
Dr. Cleotilde Gonzalez is a Research Professor with the Social and Decision Sciences Department and the Founding Director of the Dynamic Decision Making Laboratory at Carnegie Mellon University. She has additional affiliations with the Human-Computer Interaction Institute, the Societal Computing program, and the Security and Privacy Institute at CMU. She is a lifetime Fellow of the Cognitive Science Society and the Human Factors and Ergonomics Society. She is also a member of the Governing Board of the Cognitive Science Society. She is a Senior Editor for Topics in Cognitive Science, a Consulting Editor for Decision, and Associate Editor for the System Dynamics Review. She is also a member of editorial boards in journals including: Cognitive Science, Journal of Experimental Psychology-General, Human Factors. She is widely published across many fields deriving from her contributions to Cognitive Science and her computational modeling work including the development of a theory of decisions from experience called Instance-Based Learning Theory (IBLT). She has been Principal or Co-Investigator on a wide range of multi-million and multi-year collaborative efforts with government and industry, including current efforts on Collaborative Research Alliances and Multi-University Research Initiative grants from the Army Research Laboratories and Army Research Office. She has mentored more than 30 post-doctoral fellows and doctoral students, many of whom have pursued successful careers in academia, government, and industry.
Day 2: Wednesday May 31, 2023
Session 4
Panel 4 Title: Multiple Models - Role of Artificial Intelligence and Cyberinfrastructure
Speaker 1: Dr. Arjan Durresi
Title: Using Trust to Integrate and Control Multiple Models in Decision Making
Abstract
In this talk, we will discuss examples of our work on using our trust framework to integrate multiple models in decision-making in various fields of applications. Furthermore, our approaches can be used successfully as a feedback system to control algorithms, reduce uncertainty and align their results.
Biography
Arjan Durresi is a Professor of Computer Science at Indiana University Purdue University in Indianapolis, Indiana. He previously held positions at Ohio State University and LSU. His research interests include Trust Engineering, Trustworthy Artificial Intelligence, AI Control, Security, Network Architectures and Protocols, and Quantum Computing.
He has published over 100 papers in journals, over 220 articles in conference proceedings, and twelve book chapters in these fields. And he was named among the top 2% of scientists on Stanford's list in September 2021 and updated in October 2022.
Speaker 2: Dr. Sudhakar Pamidighantam
Title: FutureWater: Building Community Cyberinfrastructure for Modeling Water Resources in Indiana Under a Changing Climate
Abstract
Biography
Dr. Sudhakar Pamidighantam has been developing and deploying production cyberinfrastructures for various disciplines such as chemistry and physics supported by NSF. He is a project management committee (PMC) member of the Apache Airavata Science Gateway middleware framework. He deployed Chemviz, the chemistry educational portal with integration of NCSA Condor resources during 90s and developed GridChem cyber infrastructure during 2005 and his most recent cyberinfrastructure project is the Science and Engineering Applications Grid (SEAGrid, SEAGrid.org). His research involves Ab initio and molecular dynamics modeling and reaction mechanism, and cyber infrastructure for complex workflows. His interests are in providing production quality e-infrastructures for multi-disciplinary research using scientific and engineering modeling and simulation workflows. Sudhakar Pamidighantam obtained his B. Sc. From Madras University and M.Sc. from University of Hyderabad. After spending couple of years at Indian Institute of Science, Bangalore as a UGC Junior Research Fellow he moved to University of Alabama at Birmingham to complete his Ph. D in the department of chemistry. He joined Indiana University in 2014 as a senior scientist in the Science Gateways Research Center in the Pervasive Technology Institute. He has been a consulting and affiliate research scientist for high performance computing and material science applications at the National Center for Supercomputing Applications (NCSA) at the University of Illinois since the last 30 years. He is an affiliate of the IEEE and member of the American Chemical Society, American Association of Advancement of Sciences and the Association of Computing Machinery.
For more information, check the
Workshop flier Pdf
Workshop Acknowledgements: National Science Foundation (Award # 2228765)