The Method of Anchored Distributions (MAD): Principles and Implementation as a Community Resource

For a listing of MAD-related publications, see the Documentation page. 

In 2010, we proposed to develop a community-based, open-source computational platform for Bayesian characterization of uncertain parameter fields. It was motivated by the need to provide easily-accessible tools for applications and a platform for broad-based, long-term development, built to meet the challenges brought upon by the ever increasing complexity of modern computational tools, the diversity of data types and the breadth and depth of the subject matters needed for applications. The initiation of such an effort was intended to emulate the success experienced by other communities with a long tradition of community-based development efforts, and our hope was that it could lead to a transformative change in hydrology. The absence of such an approach had been hurting the science discovery process. 

Our project was to construct a modular, open-source computational platform for application and testing of the general statistical framework for parameter field characterization.  Our approach was built around two concepts that make it applicable to a wide range of data types and computational tools. The first concept is a classification of data into two broad categories. The second concept is a strategy for localization of data using anchors (hence Method of Anchored Distributions, or MAD) that captures the relevant information contained in a local or non-local process, respectively, and presents it in terms relevant for the parameter field(s). 

We have constructed a modular computational platform, with 2 distinct modules, which deliver the capabilities necessary to support the aforementioned concepts: namely, a pre-processing block and a post-processing block. The modules comprise the MAD kernel. The design of the user-interface allows users and developers to easily link the kernel with their models via a driver. Many of the decisions over assumptions to be adopted or computational tools to be employed are at the user’s discretion and are not hard-wired into the kernel. The kernel is built using open-source architecture.

Eventually, two versions of MAD will be available: a super-computing version (for large-scale applications - in development) and a desktop application (for educational purposes and smaller-scale applications). Also, the kernel will be eventually interoperable with a wider  array of community-based resources (such as CUAHSI’s hydrologic information systems) and is already interoperable with some libraries (such as R). With this framework for generic application of the concepts and principles of MAD in place, further types of data can be introduced, simplifying assumptions can be added or removed at the user’s discretion, and powerful numerical models and analytical tools, current or under development, can be integrated much more easily in future analyses under long-term community-based development to follow. 

 

Last edited May 1, 2016 at 8:10 PM by frystacka, version 23