One of the tools for determining nonlinear effects and interactions between the explanatory variables in a logistic regression model is using of evolutionary product unit neural networks. To estimate the model parameters constructed by this method, a combination of evolutionary algorithms and classical optimization tools is used. In this paper, we change the structure of neural networks in the form that all model parameters can be estimated by using an evolutionary algorithms causes a model that is Akaike information criterion is better than conventional logisti model Akaike information criterion, but using the combination method gives the best model.

Consider an experimental situation where it is desired to compare more than one test treatments with a control treatment. In this paper a method is presented for achieving E-optimal incomplete block design for this situation under the assumption that the observations within each block are correlated. Then an algorithm is provided for making optimal design based on above-mentioned method. This algorithm for any correlation structure with negative non-diagonal elements is applicable.

Logistic regression models in classification problems by assuming the linear effects of covariates is a modeling for class membership posterior probabilities. The main problem that includes nonlinear combinations of covariates is maximum likelihood estimation (MLE) of the model parameters. In recent investigations, an approach of solving this problem is combination of neural networks, evolutionary algorithms and MLE methods. In this paper, another type of radial basis functions, namely inverse multiquadratic functions and hybrid method, are considered for estimating the parameters of these models. The experimental results of comparing the proposed models show that the inverse multiquadratic functions compared to the Gaussian functions have better precision in classification problems.

Neighbour effects, that is the response on a given plot is affected by the treatments in neighbouring plot and the effect by the treatment applied to that plot. As a result, the estimate of treatment differences may deviate because of this interference from neighbouring plots. Neighbour-balanced designs ensure that the treatment comparisons will be as little affected by neighbour effects as possible. Circular neighbour-balanced design are divided into two groups. In the previouse researchs, method of cyclic shifts to construct CNB1 has been used, the authors used this method to construct CNB2. Some series of CNB2 are found by omputer programming using in MATLAB software and method of cyclic shifts. Then, some of these designs witch are universally optimal under models with one sided neighbour effect (M1) are identified.