MMA, A Computer Code for Multi-Model Analysis

Abstract

The Multi-Model Analysis (MMA) computer code is designed to evaluate many alternative models of a given system, called multiple models in this work. It can be used to rank the models and calculate posterior model probabilities. The probabilities are used to calculate modelaveraged quantities that account for the variability evident in the alternative models. The modelaveraged quantities can include parameter estimates, predictions, and measures of parameter and prediction uncertainty. Calibration of all models needs to be completed before application of MMA. MMA operates by reading files produced by models developed and calibrated to represent a single system. The models all need to use the same observations with the same weighting. The files needed by MMA can be produced by UCODE_2005 and associated codes (Poeter and others, 2005) run in the Sensitivity-Analysis or Parameter-Estimation mode. The name and structure of the files are simple and follow the conventions of JUPITER API data-exchange files. Thus, it is likely that MMA also will be able to use results from codes other than UCODE_2005. Although the examples presented in this work are from the field of groundwater modeling, MMA (and UCODE_2005) can be used to evaluate results from models of nearly any type of system. The documentation presented in this report describes the methods used and the input and output files. Together, this report, Hill and Tiedeman (2007), and Burnham and Anderson (2002) serve to document the computer code described in this work. Additional information on the methods is provided in Burnham and Anderson (2004), Ye and others (2004, 2005), and Poeter and Anderson (2005). The performance of MMA has been tested in a variety of applications. Future applications, however, might reveal errors that were not detected in the test simulations. Users are requested to notify the originating office of any errors found in the report or the computer program. Updates might occasionally be made to both the report and the computer program. Users can check for updates on the World Wide Web at URL

Multiple Parameterization for Hydraulic Conductivity Identification

“…Several information criteria other than BIC, for example, the Akaike information criterion (AIC), corrected Akaike information criterion (AICc), and Hannan and Quinn’s information criterion (HQ), have been introduced to estimate the model weights in (Poeter and Anderson 2005; Peoter and Hill 2007). The major difference between BIC and AIC, AICc, and HQ is the penalty term (the second term in ), which penalizes the number of unknown parameters in lieu of the parsimony principle.…”

A review of regional groundwater flow modeling

“…Assessment of alternative conceptual models has received increased attention (Neuman, 2003;Ye et al, 2004;Hojberg and Refsgaard, 2005;Poeter and Anderson, 2005;Refsgaard et al, 2006Refsgaard et al, , 2007. A general purpose computer software for multi-model analysis has recently been developed (Poeter and Hill, 2007). These methods have been applied to the northern Yucca Flat area of the Nevada Test Site located within the Death Valley Regional Flow System (Ye et al, 2010).…”

Programs for Calibration‐Based Monte Carlo Simulation of Recharge Areas

“…Either of these methods could be run in concert with Monte Carlo simulation by including them in the flow model. This method, like most other methods, assumes that the structure of the model is correct, that is, the model geometry and boundaries are sufficient to describe the behavior of the system. If alternative model structures are thought to be possible, the user should consider multimodel analysis (Poeter and Hill 2007). The method is intended to simulate contributing areas using MODPATH where particles are captured by a sink.…”

“…This method, like most other methods, assumes that the structure of the model is correct, that is, the model geometry and boundaries are sufficient to describe the behavior of the system. If alternative model structures are thought to be possible, the user should consider multimodel analysis (Poeter and Hill 2007).…”