This paper proposes a method for Soft Sensors design using a Multilayer Perceptron model based on co-evolutionary genetic algorithms, called CEV-MLP. This method jointly and automatically selects the best input variables and the best configuration of the network for the prediction setting. The CEV-MLP is constituted by three levels, the first level selects the best input variables and respective delays set, the second level is composed by the parameters of hidden layers to be optimized (number of neurons in the hidden layers and transfer function), and the third level is the combination of first and second level. The method was successfully applied, and compared with two state-of-the-art methods, in three real datasets. In all the experiments, the proposed method shows the best approximation accuracy, while all the design of the prediction setting is performed automatically.