The goal of the work is to investigate and compare the effectiveness of four deep learning optimization algorithms SGD, Adam, Adagrad and RMSprop in forecasting and pattern recognition tasks. The results of the work – in the process of the work, a comparative analysis of deep learning optimization algorithms was carried out. For this, 5 data sets were selected and prepared for model training. Models with different architectures were created according to each data set. Models were compiled and trained. At the compilation stage, the loss function for each model was determined according to the task, as well as the metrics by which the performance of the model will be evaluated. After training the model, the corresponding losses and metrics were measured on the test data. All results were analyzed and conclusions drawn.