Designing for Change Simulation Adaptation Composition and Reuse

Designing for Change: Simulation Adaptation, Composition and Reuse Paul F. Reynolds, Jr. reynolds@virginia. edu 434 924 -1039 David Brogan, Rob Bartholet, Joe Carnahan, Xinyu Liu, Ross Gore, Lingjia Tang, Yannick Loitière, Michael Spiegel, Chris White Modeling and Simulation Technology Research Initiative Computer Science Department University of Virginia, USA http: //www. cs. virginia. edu/~Ma. STRI/

What Are Key M&S Challenges? e s u e r simulation Composition – Constructing ar (larger) from two or more existingfo simulations. g Interoperability – When two or more simulations in can execute together meaningfully. t for a purpose other than Reuse – Using ap simulation ait was originally intended. that for which d A Multi-resolution modeling – Simulating overlapping phenomena at two or more levels of (spatial/temporal) resolution concurrently.

Modeling and Simulation Technology Research Initiative Simulation design and semi-automated adaptation for reuse • COERCE – Coercibility: the practices and methods for capturing designer knowledge in software – Coercion: a user-guided, semi-automated software adaptation process • Composability – Reusing components, possibly with acceptable amounts of revision, to meet new requirements

Coercibility Expansion Opportunities Design-time provided insights Sensitivity Analyses Flexible Points Simulation Metadata for coercing

Coercion metadata New requirements Expansion Opportunities Sensitivity Analyses Flexible Points Simulation e Coercion vis iz al u Modify? (solid insight) Optimize? (need to explore)

coerce(model) ® what-I-want

Full physics bicyclist 2 D point mass

Collaborators at UVa • Aerospace engineering (thermo-acoustic coupling) • High-energy physics (quark-gluon plasma) • Neuroscience (multi-scale model of hippocampus) • Earth Science (forest respiration) • Biomedical Science (proposal: cell/tissue MRM)

Outline of talk Coercibility Composability Coercion

Coercibility Capturing and encoding the nature of alternative conditions (not specifics) • Stipulate when simulation can/cannot be used • State when/how it can be changed • Identify critical dependencies

Designing for change Build upon simplifying assumptions • Frequently hidden • Small changes can invalidate simulation • Examples: – Bounding the space and time of the simulation – Selecting equations to represent phenomena – Knowing 4 th order Runge Kutta is good enough

Case study Defining a simulation’s context • Articulate the assumptions in Falling Body Model – A sphere falls to earth… – 10 people listed assumptions – 29 assumptions were ultimately identified § The top three people only found 21, 19, and 16 Spiegel et al. , WSC 2005

Capturing Insight Flexible Points • Any element of the model or simulation that can be manipulated in meaningful and effective ways to direct a simulation’s behavior • Examples: – Value substitution for constants/parameters – Replacing abstraction assumptions – Modifying stochastic elements – Tuning logical time components

Clarifying Flexible Points Pragmatic definition, emphasis on semiautomated transformation Parameters Contrast with Flexible points • Simulation parameters • Design decisions Carnahan et al. Fall SIW 2004 Carnahan et al. , WSC 2005

Composability “Assembly of parts into a whole without modifying the parts. ” (Szyperski 2002) “The capability to select and assemble simulation components in various combinations into valid simulation systems to satisfy specific user requirements. ” (Petty and Weisel 2003) We advocate semi-automated adaptation of components to support composability. --practical and realistic

Composability It’s hard to build component repositories • Size of components • Component interfaces • Classifying semantics It’s hard to build compositions • Some theoretical results, few practical

Component Selection (CS) REQUIREMENTS R COMPONENTS X x 1 r 4 r 1 x 2 x 4 r 2 x 5 r 3 r 5 r 6 r 7 x 3 r 8 x 7 x 6 x 8 CS: Is there a subset of X of cardinality k or less that covers R? Example instance when k=3

Component Selection CS is NP-complete • Reduction from SAT (Page and Opper 1999) • Reduction from MSC (Petty et al. 2003) Approximate solutions to CS are possible • Limit the amount of emergence • Use Greedy algorithm Fox et al. WSC 2004

Applied Simulation Component Reuse (ASCR) Critical New Assumption: Any component can be adapted to satisfy any requirement What does this buy us? • We no longer have to assume the existence of a master set of components. • We can more flexibly react to changing requirements. But… We now have to account for the cost or utility of adapting a component!

ASCR A formal model and analysis of component selection with adaptable components • What is utility and cost of adapting component to satisfy additional requirements? • What is value of incomplete satisfaction?

Results Proven: Exact cover by three sets (X 3 C) reduces to ASCR • brings flexibility to component selection, but the problem remains intractable (NP-Hard) Ongoing work: • Discovering scalable methods, algorithms, and heuristics for component selection • Encoding adaptability into the component Bartholet et al. , WSC 2005

Coercion Efficient Adaptation of a simulation for reuse and component selection • Flexible points • Language tools • Automatic visualization • Sensitivity Analysis • Optimization Support user seeking insight

Optimization Traditionally used to find best parameter values Generates additional insight • Identify sensitive/brittle systems • Explore novel circumstances • Detect correlations • Discover constraints • Bound search space

P: The Coercion Process S 0 ? S? S? . . Btarget

P: The Coercion Process S 0 Satisfies Btarget? P Yes No Run optimizer or modify? Optimize Modify Sn Btarget

Paths to Coercion S 0 P Sn S?

Dangerous Divergence S 0 Sn

Distance Function S 0 Si I i Insight is vital! Btarget Distance Function D(Si, Ii) Coercion converges on solution when D(Sn, In)==0

The Convergence of Coercion How can we guarantee coercion will terminate? • We benefit from our successes – Better satisfy Btarget with good transformations • We learn from our mistakes – Gain insight from bad transformations

Focus on User Interaction with Optimization Tools Explore what you don’t know and exploit what you do know Simulated Annealing Explore Genetic Algorithms Unknown Gradient-Based Search Random Code Generation Response Surface Methodology Code Modification Exploit Insight

Running Example Global Minimum

Simulated Annealing Global Minimum

Simulated Annealing Insight: Explore Unknown Simulated Annealing • Perturbation func. • Cooling schedule Exploit Insight

Genetic Algorithms

Genetic Algorithms Insight: Explore Unknown • Mutation func. Genetic Algorithms • Crossover func. Exploit Insight

Gradient-Based Search

Gradient-Based Search Insight: Explore Unknown • Initial guess Gradient-Based Search Exploit Insight

Response Surface Methodology

Response Surface Methodology Insight: Explore Unknown • What to inspect • Where to inspect • What to infer Response Surface Methodology Exploit Insight

Optimization Techniques Explore Unknown Simulated Annealing Genetic Algorithms Response Surface Methodology Gradient-Based Search Exploit Insight Waziruddin et al. Fall SIW 2004

Identify/classify best practices for optimization in coercion Set-up time Computation time Technique preemption

Set-up Time Explore Unknown More Set-up Time Simulated Annealing Genetic Algorithms Response Surface Methodology Gradient-Based Search Exploit Insight

Computation Time More Computation Time Explore Unknown Simulated Annealing Genetic Algorithms Response Surface Methodology Gradient-Based Search Exploit Insight

Technique Preemption More Preemption Explore Unknown Simulated Annealing Genetic Algorithms Response Surface Methodology Gradient-Based Search Exploit Insight

Summary • Efficient simulation adaptation appears viable for simulation reuse • Currently funded under NSF (ITR) DDDAS program: dynamic adaptation • Quite interested in collaborations with application experts. reynolds@cs. virginia. edu dbrogan@cs. virginia. edu