Abstract. In order to overcome the illposedness of groundwater inverse analysis it is inevitable to introduce prior information of some form and thus Bayesian statistics. One of the essential problems in Bayesian inverse formulation is the optimum matching between the objective information (i.e., the observation) and the subjective information (i.e., the prior information). In this study, Akaike's Bayesian Information Criterion (ABIC) is introduced to overcome this problem. ABIC is also effective in the model identification problem, and this aspect is also emphasized. The effectiveness of the method is illustrated by analyses on an actual aquifer system. Both steady and transient state analyses are carried out. The paper also provides the background of ABIC in some detail. The method also recognizes the dour fact that whether the prior statistics are determined from actual field data or borrowed from other aquifers, they will seldom be known with sufficient accuracy to justify imposing them on the inverse solution without further modification; for this reason we do not require that the covariance matrices of the random vectors entering into the model be specified exactly, but to a scaler of multiplication.
Model Selection and ParameterizationThe other important aspect of groundwater inverse analysis, besides the matching of the objective and subjective information, is the model selection or model identification problem. It is well recognized in the multiple regression analysis that if one increases the number of explanatory variables in a regression model, fit of the model to the given data improves, whereas reliability of estimated parameters decreases, and vice versa. There is a tradeoff between the parameter dimension and the reliability of estimation for estimated parameter values. In groundwater inverse analysis this problem typically appears in selecting number of zones in calculation model. This problem has been studied by many people; an excellent review on the topic is given by Yeh [1986]. Correra and Neuman [1986a, b, and c] solved this p•/oblem by using Akaike's Information Criterion (AIC). From the start of the development, AIC was aiming at solving this problem of model selection; it worked most conveniently for this purpose. Akaike's Bayesian Information Criterion (ABIC) introduced in this study is also developed on the same information theory principal as AIC; it can also be used to select the best model from several alternative models as is illustrated in the example in this paper.