Model order selection

It is also important to realize that a fifth-order state space model has considerably more than five model parameters  

One often speaks in this context about the curse of dimensionality, by making the model slightly more complex, one increases the number of model parameters significantly. One should therefore watch out for models that produce a good model fit due to a high model dimensionality  

---

Several asymptotic model order selection methods use a form of generalized information criteria (GIC) that could be represented as GIC(a, p) = N In(pp) + ap, where p is the model order, o is a constant, N is the number of data points, and Pp is the variance in the residual or error for model p. The error or residual variance pp can be determined using a (forward) prediction error p(n) defined as... 

Prom the results presented in Table 3.1, our conclusion is that the best ARX model, in terms of predictive capability, for the Process Trainer is either a 14th order model (28 model terms) with d = 1 or a. IZth order model (26 model terms) with d — 2. The reason such high order models were selected is that, with low order ARX models, there must exist a mismatch between the assumed and actual noise structures. Evidence for this statement can be found in Ljung (1987) where the addition of a noise model was found to give improvement in terms of the AIC. In a similar situation, Kosut and Anderson (1994) have fit least squares ARX models using cross validation for model order selection and have found that high order ARX models are often necessary. 

This suggests that the errors in the estimated high frequency parameters are the reason for the lads of smoothness in the step response estimates. To confirm this conjecture, we estimate step response models using various reduced order FSF models. The model orders selected aure n = 99, 49, 25 and 11. Figure 5.11 shows the estimated step response models for these four choices of n where it can be seen that, as more high frequency parameters are deleted from the estimated FSF model, the step response model becomes smoother. Also, as the number of estimated high frequency parameters is decreased, the numerical conditioning of the FSF correlation matrix improves. When n = 49, the condition number is 1684.7, with n = 25 the... 

For example, Figure 5.18 shows the estimated step responses for u to j/2-The FSF model order selected for this input-output paur was 3 which means that the true response may not lie within the confidence bounds. Evidence for this is found by observing that the confidence boimds at time zero do not include zero as a possible value. However, for most processes encountered in the process industries, the initial value of the step response is known to be equal to zero. By comparison, for higher order models, the confidence bounds seem to always enclose zero as a possible value at time zero. For example, Figure 5.19 shows the estimated step responses for ui to yi that has a selected model order of 15. In this case, the bounds aure believed to enclose the true response with a 99% confidence level. 

Goodwin, G. C., M. Gevers B. Ninness (1992), Quantifying the error in estimated transfer functions with application to model order selection , IEEE Transactions on Automatic Control 37, 913-928. 

Phase I. AR/MA orders selection. In order to decouple the selection of the model orders tta, tic) from that of functional subspaces, their interaction has to be minimized. For this reason a high-degree P and the complete set of PC basis functions may be initially adopted. When employing these, AR/MA model order selection may be achieved through trial-and-error techniques based on the values of the fitness function. 

By structural complementarity, dicationic l,4-diazabicyclo[2.2.2]octane (VII) provides an appropriate recognition site for phosphate ions and two stearyl side chains attached to the amines add lipophilic properties 59,60). Such a carrier model can selectively extract nucleotides from aqueous solution to chloroform solution via lipophilic salt formation. The order of nucleotide affinity is ATP > ADP > AMP. The selectivity ratios were 45 for ADP/AMP and 7500 for ATP/AMP at pH 3. The relative transport rate was ATP > ADP > AMP. The ratios were 60 for ATP/AMP and 51 for ADP/AMP. The modes of interaction of ADP and ATP are proposed to be as shown in Fig. 6. 

Under this project, an IPCS Harmonization Project Document on the Principles of Characterizing and Applying Human Exposure has been published (WHO/IPCS 2005). This document sets out the characteristics of exposure assessment models that should be described to aid in model selection by exposure assessors. The document summarizes current practice in exposure modeling and principles for evaluating exposure models, but does not provide a comprehensive list of existing exposure models. The focus of the document is on the discussion of general properties of exposure models and how they should be described. The characteristics of different modeling frameworks are examined, and 10 principles are recommended for characterization, evaluation, and use of exposure models in order to help model users select and apply the most appropriate models. The report also discusses issues such as validation, input data needs, time resolution, and extrapolation of the model results to different populations and scenarios. 

Repeat the input identification experiment with the model order MD = 2. Compare the linear regression residual errors for the two cases. Select the "best" model order on the basis of the Akaike Information Criterion (see Section 3.10.3 and ref. 27). ... 

Typically the models have horizontal resolutions of the order of 3-6 degrees, with the exception of one of the models which has a horizontal resolution of 8 degrees x 10 degrees. In the vertical, the models have from 9 to 25 levels. There is a wide variety in vertical domain, with the top layer ranging from 100 to 10 hPa. For some of the models the selection of upper boundary therefore is a limitation in the models ability to predict aircraft impact, which mainly occur in the upper troposphere and lower stratosphere. 

There are several methods that can be used to select well-distributed calibration samples from a set of such happenstance data. One simple method, called leverage-based selection, is to run a PCA analysis on the calibration data, and select a subset of calibration samples that have extreme values of the leverage for each of the significant PCs in the model. The selected samples will be those that have extreme responses in their analytical profiles. In order to cover the sample states better, it would also be wise to add samples that have low leverage values for each of the PCs, so that the center samples with more normal analytical responses are well represented as well. Otherwise, it would be very difficult for the predictive model to characterize any non-linear response effects in the analytical data. In PAC, where spectroscopy and chromatography methods are common, it is better to assume that non-linear effects in the analytical responses could be present than to assume that they are not. 

Dougal RA, Gao L, Liu S. Ultracapacitor model with automatic order selection and capacity for dynamic system simulation. Journal of Power Sources 2004 126 250-257. 

Figure 5. Actual human development index (HDI) values versus their cross-validation predictions from regression models (a) selected from all first-order terms (b) selected from all second-order terms.

An electric motor of efficiency 0.92 and a manual 1-speed transmission have been used for all the simulations. The transmission often constitutes a significant loss factor, and losses in using 5-speed or automatic transmissions (Cuddy, 1998) are higher than for the 1-speed gearbox. Because the actual Lupo 3L has n electronically operated automatic transmission with higher efficiency than the corresponding manual transmission, the 1-speed model was selected in order to avoid a distorting impact on the simulation results. 

T.-J. Wu and A. Sepulveda, "The Weighted Average Information Criterion for Order Selection in Time Series and Regression Models," Statistics Probability Letters, 39 (1998) 1-10. 

Model-based approaches allow fast derivative computation by relying on a process model, yet only approximate derivatives are obtained. In self-optimizing control [12,21], the idea is to use a plant model to select linear combinations of outputs, the tracking of which results in optimal performance, also in the presence of uncertainty in other words, these linear combinations of outputs approximate the process derivatives. Also, a way of calculating the gradient based on the theory of neighbouring extremals has been presented in [13] however, an important limitation of this approach is that it provides only a first-order approximation and that the accuracy of the derivatives depends strongly on the reliability of the plant model. 

---