A supervised feedforward artificial neural network (ANN) trained with backpropagation algorithms, along with multiple linear regression (MLR) model, was applied to predict the biological yield (BY) of grass pea. For this purpose, a four-year study (2008-2012) was carried out under rainfed conditions in three dryland stations in Iran, and a panel of 12 grass pea genotypes were evaluated for agronomic and phonologic characters. Stepwise regression (SWR) and principal component analysis (PCA) were employed to evaluate 15 input parameters. To discover the optimum ANN structure, different training algorithms, transfer functions, number of hidden layers and neuron in each layer were studied and optimized using Taguchi's method. The multilayer perceptron (MLP) model with tangent sigmoid (tagsig) transfer function, Levenberg-Marquardt learning algorithm, and two hidden layers was identified as the best model to predict the BY of the grass pea. The sensitivity analysis of inputs revealed that seed yield (SY) followed by thousand seed weight (TSW) and number of days to maturity (NDM), respectively, were the most influential factors in predicting BY in both models. According to the adjusted ANN model, early flowering genotypes with long maturity and high TSW should be considered as the best model for enhancing BY in breeding programs.