From Maps to Models Augmenting the Nation's Geospatial Intelligence Capabilities (2016) / Chapter Skim
Currently Skimming:
3 Model Foundations
3 Model Foundations
Pages 37-56
The Chapter Skim interface presents what we've algorithmically identified as the most significant single chunk of text within every page in the chapter.
Select key terms on the right to highlight them within pages of the chapter.
From page 37... ...
could think about them from a practical standpoint. The first section focuses on models, including whether to develop new models or use existing models and what types of models to use in the investigation.
|
From page 38... ...
However, because they were developed for a different purpose, existing models or model output will likely need some adaptation or augmentation for the investigation at hand. Advantages and disadvantages of developing models, combining existing or new models, using existing model output, and running an existing model are outlined below.
|
From page 39... ...
Examples include geospatial data, social media data, specialized spatial data or geo network models, or output from a social system model evolving in a spatial environment. If the system exhibits spatial or network dependence, standard statistical approaches that rely on independence of the observations and entities will not be appropriate for analysis.
|
From page 40... ...
Developing process models is typically demanding and involved. If many components and feedbacks are required to represent the system, model development will require substantial labor, time, and computing resources.
|
From page 41... ...
A key consideration is how strongly the subsystem models being combined are expected to couple and interact with one another. The pirate example in Chapter 2 illustrates one-way coupling, because outputs of the weather and wave model serve as input to the pirate behavior model, but the pirate activity does not feed back to the weather model.
|
From page 42... ...
has utilities for managing many model runs on a variety of computing systems, as well as a collection of analysis capabilities to process the resulting model output. Existing models can be run using an iterative approach, in which the results of the model run (or ensemble of model runs)
|
From page 43... ...
. Iterative methods have proven effective for successive computational model runs that can be carried out relatively quickly (e.g., fractions of a second, or minutes)
|
From page 44... ...
Often modeling approaches, analysis methods, and computational infrastructure need to be developed in concert to efficiently leverage the data being collected for the investigation. Ingesting geospatial information into models raises special considerations.
|
From page 45... ...
How will verification, validation, and uncertainty quantification be carried out to support model assessment? How model assessment tasks are carried out depends on the properties of the model, the availability of relevant data, and the nature of the key questions in the investigation.
|
From page 46... ...
In more exploratory investigations, uncertainty quantification might involve a sensitivity analysis, exploring a range of possible model outcomes to understand the effect of input and parameter uncertainty. With sensitivity analyses, it is important to consider how changing multiple inputs simultaneously affects the resulting model output (Saltelli et al., 2008)
|
From page 47... ...
For example, while each of the approximately 20 climate models in the Coupled Model Intercomparison Project (CMIP) had its own biases, the prediction accuracy of the mean across all the models was better than that of any individual model (Reichler and Kim, 2008; see Figure 3.6)
|
From page 48... ...
Large-scale computational models, such as the weather models mentioned above, are designed to run on compute-intensive high-performance computing systems for hours or days, running on thousands of processors, holding huge state vectors in distributed memory and storage, making use of specialized graphical processing units, and producing high volumes of model output that require specialized data storage capability (see Appendix B)
|
From page 49... ...
Data-Intensive Computing The modern deluge of data (e.g., from social media, automated transaction records, remotely sensed data, scanner data, text, and computational model output) has motivated the conception and expansion of data-intensive computing.
|
From page 50... ...
The MapReduce paradigm carries out operations on data residing on a distributed file system by distributing a task over the data set, carrying out the tasks locally on the separate data pieces, collecting the distributed intermediate results, then producing the final result. This approach is efficient for embarrassingly parallel tasks,
|
From page 51... ...
Spatial Computing Spatial computing covers computing in spatial, temporal, and spatiotemporal spaces. Models and data with spatial or space-time-dependent structures do not fit easily into either data-intensive or traditional high-performance computing-based analysis frameworks for a number of reasons, such as the following: • It is difficult to divide many spatial analysis tasks into equal subtasks that impose comparable costs to different processors because of spatial data diversity and spatial variability in data density.
|
From page 52... ...
For example, climate models that use cutting-edge high-performance computing can run at higher resolution or include more detail in the modeling, perhaps yielding results that are more like the real-world system of interest. Similarly, analysis approaches that use data-intensive computing can handle a much larger data volume and rate for constructing empirical models or carrying out inference, potentially producing faster and more comprehensive results.
|
From page 53... ...
• Existing models • Are the key questions "extrapolative" or captured in past data and experience? • Data • Amenability of the real-world system to modeling to address key questions • Computation • Accuracy of the result required A key property of a model's impact on the investigation is its fidelity to the real-world system being modeled.
|
From page 54... ...
A model based on a simple, quick approach may be more useful than one that is more comprehensive but slow. In such cases, it may be possible to utilize existing models or model output, which would both speed the investigation and reduce its cost.
|
From page 55... ...
The computational infrastructure needed depends on the type of model being run. In general, large-scale, physical process models use traditional high-performance computing for models that require large numbers of processors, fast communication between processors, and large volumes of data in memory and storage (e.g., climate models)
|
Key Terms
This material may be derived from roughly machine-read images, and so is provided only to facilitate research.
More
information on Chapter Skim is available.