Ruba Aljafari speaks tonight (with contributions from Drs. Deokar and El-Gayar) on model management.

Dr. Deokar notes that their modeling work is related to semantic web technologies; he’s teaching a course on that this summer (I’m registered).

Model Management:

  • centers on study of computer-based methods for representing models and automating modeling processes
  • emerged as result of shifting emphasis of DSS from data and data analysis to decision models
  • big review: read Krishnan and Chari (2000), “Model Management: Survey, Future Research Directions, and Bibliography,” Interactive Transactions of OP/MS….
  • supports some tasks in complex modeling cycle; some tasks still require human intervention (like on problem identification)
  • MM functionalities resemble database management systems (DBMS) such as model description, manipulation, control, and selection
  • operations of relational data management can be extended to include models
    • two models can be joined when output of one can be used as input to another
    • languages similar to SQL and QBE have been proposed for relational model bases
  • model processing focuses on issues such as interfacing models with users….

Model selection: key issues:

  • organizational
    • should models be grouped by similarity?
    • which grouping can be actually meaningful to the modeler?
  • representation
    • what features play role in model selection?
    • how design representation to accommodate new models that will be added to the library?
  • processing
    • what set of ops useful to modeler?
    • what should be the expressive power of a query language?…

Model selection is simple compared to other functionalities; it’s just retrieval! Model composition is hard:

  • main purpose: link interdependent models in specific cases where no one model fully addresses the problem
  • how should links (between models) be represented?
  • how should dynamic variable linkages be supported? when is a link well-formed?
  • how to declare ruels to determine well-formedness of links?
  • how represent info used to make that determination?
  • at what level of abstraction should model be built?
  • can composition procedure be automated to minimize human intervention?…
  • can order of models be inferred?…
  • how can model composition and execution be realized in a distributed heterogeneous environment?

Model Integration: very similar to composition, but more complex:

  • involves schema integration and process (solver) integration
  • composition just deals with linking via inputs and outputs
  • integration actually modifies the structure of existing models
  • resolving conflicts is a big deal (variables with same name but representing diff elements in diff models)
  • interaction effects mess things up
  • black box vs. glass box? black box model doesn’t show modeler the structure; glass box does (may relate to privacy issues, proprietary knowledge: not every business model is going to tell you exactly what factors are used; think secret sauce or BNN influence ratings? strategic partners may not want to share everything… oh my!)

Model Formulation: converting problem into mathematical model

  • must figure out where MM tech can be of most use
  • what form should inputs take?
  • reasoning issues: how to turn qualitative description into formula?

Model Implementation: creating a model representation to which a solver can be applied

  • main principles include model data independence (gotta be able to use it on other data sets), solver independence, problem paradigm independence, and meta-level representation (gotta be able to represent info about the model)
  • algebraic modeling languages: GAMS, LINGO
  • structured modeling languages: SML, LSM, “structured modeling markup language”

Our main focus: Model Representation

Three levels of abstraction:

  1. fundamental constructs and relationships among them to describe a particular modeling paradigm (e.g. supplier, customer, distrib. center)
  2. particular model schema indep. of its data (e.g. structure of transportation problem)
  3. particular model instance (e.g. specific transport problem fed with specific dataset, specific customers and suppliers)

Structured modeling is a couple decades old, but advancing to distributed environments requires finding way to share the models. XML has emerged as a format amenable to distributed environments; thus logical to convert modeling paradigms into XML formats.

Motivation behind structured modeling: low productivity of management sciences/operations research (in other words, they don’t get the job done or produce spectacular results for the investment) and poor managerial acceptance (manager looks at LINGO model and says, “What the heck is this?” Go get a degree, and it will make perfect sense, but if you have work to do, it’s not going to speak to you.)

What can structured modeling offer?

  • single framework for representing models, suitable for computer execution, managerial communication, and mathematical use
  • independence of rep and solution
  • suffficient generalizability
  • usefulness for most of life cycle
  • desktop implementation
  • integrated facilities for data management
  • immediate expression evaluation…

Structured Modeling framework (not a language!) has three levels

  1. elemental: captures all definitional detail of specific model instance (five types: primitive entity [supplier: can stand alone], compound entity [link: can’t stand alone!], attribute, function [profit: calculated or derived value], test[true/false! boolean!])
  2. generic: captures natural familial groupings of elements
  3. modular: organizes generic structure hierarchically

SM can have static and dynamic models; dynamic possibly by allowing attributes and functions to be functions of time

Model schema: “any class of structured models whose modular outlines all can be placed in a 1:1 relationship in a way that is consistent with modular structure, generic structure, and the intended meaning of the models”


DSU Research on Model Management

Goal is to facilitate reuse of models through discovery, sharing, and composition. Remember, this can be KM stuff! models can encapsulate important knowledge, help you design, deliver, and manage services….

Research Questions:

  1. How can we add semantics to decision-making models?
    1. How can we utilize semanic web tech?
    2. How can we customize these techs to fit decisoin-making models?
  2. How can we use semantics to facilitate model sharing, reuse, and composition?
    1. Can we build a prototype that can integrate the enhanced model management architecture?
    2. What kind of technologies do we need and how can we integrate them in the prototype?
  3. “What is an appropriate analytical framework  for aggregating results from multiple models or model components? Can Web services provide a technical platform for implementing model aggregation?” (Power, D.J., and Sharda, R. (2007), “Model-driven decision support systems: Concepts and research directions,” Decision Support Systems (43:3), 1044-1061)