Through employing a new error function capturing the nature of multi-instance learning, a neural network based multi-instance regression algorithm is presented in this paper. Experiments on benchmark data sets show that this algorithm deals well with the multi-instance regression problems.