Feedforward neural networks are widely used as universal predictive models to\nfit data distribution. Common gradient-based learning, however, suffers from\nmany drawbacks making the training process ineffective and time-consuming.\nAlternative randomized learning does not use gradients but selects hidden node\nparameters randomly. This makes the training process extremely fast. However,\nthe problem in randomized learning is how to determine the random parameters. A\nrecently proposed method uses autoencoders for unsupervised parameter learning.\nThis method showed superior performance on classification tasks. In this work,\nwe apply this method to regression problems, and, finding that it has some\ndrawbacks, we show how to improve it. We propose a learning method of\nautoencoders that controls the produced random weights. We also propose how to\ndetermine the biases of hidden nodes. We empirically compare autoencoder based\nlearning with other randomized learning methods proposed recently for\nregression and find that despite the proposed improvement of the autoencoder\nbased learning, it does not outperform its competitors in fitting accuracy.\nMoreover, the method is much more complex than its competitors.\n