Digital systems have been playing a vital role in various applications such as banking, finance, healthcare, manufacturing, security and so on. Also, their role and applications are becoming wider and more crucial. In many such applications, identifying and recognizing a character, or a digit accurately, plays a significant role, especially in banking and financial sectors and other sectors where an error can cause much loss or damage. This, technically, is called the optical character recognition (OCR) problem. In this context, contributionof this paper is of two folds. First, we propose a multi-layer perceptron (MLP) neural network architecture that includes an input layer, hidden layers and an output layer to develop an effective method for OCR. The architecture builds a model that learns representations from the input data and further these representations are used for classifying the unknown data. This proposed method MLP, which recognizes optical characters, is compared to existing nearest neighborhood methods such as condensed nearest neighbor (CNN),modified condensed nearest neighbor (MCNN) and other class nearest neighbor (OCNN), in performance. Posterior probabilities and conditional probabilities pertaining to recognition are computed, estimated and validated on the test data (OCR and Pendigits) for all the afore-mentioned methods. Using these posteriorprobabilities, probabilities of detection of the newly drawn character or digits can be estimated. The proposed model in this paper outperforms existing methods. The second contribution is as follows. In certain critical applications, it is very important to achieve the highest possible accuracy even if it is expensive. To achieve this a multi-pronged approach using multiple methods is developed based on these four methods, in order to improve and estimate the accuracy, in cases when multiple methods concur or otherwise.