Improve your training time, model stability and accuracy on Amazon Forecast by leveraging new hyperparameters now supported on DeepAR+
Amazon Forecast is a fully managed service that uses machine learning (ML) to generate accurate forecasts, without requiring any prior ML experience. Amazon Forecast is applicable in a wide variety of use cases, including energy demand forecasting, workforce and resource planning, cloud infrastructure usage forecasting, inventory planning, product demand forecasting, and financial planning.
Amazon Forecast is a fully managed service that uses machine learning (ML) to generate accurate forecasts, without requiring any prior ML experience. Amazon Forecast is applicable in a wide variety of use cases, including energy demand forecasting, workforce and resource planning, cloud infrastructure usage forecasting, inventory planning, product demand forecasting, and financial planning.
Today we are excited to announce support for three new DeepAR+ hyperparameters that can help reduce training time, increase model stability and accuracy. First, to improve model stability, a common issue with deep learning models where the results vary between training runs, we have introduced the hyperparameter “num_averaged_models” that allows you to average results over multiple models within a single training run. Second, to improve forecast accuracy as well as the convergence speed and thus shorten the training time, you can now change the learning rate during training with the hyperparameters “learning_rate_decay” and “max_learning_rate_decays.”
Additionally, DeepAR+ now also supports a new piecewise-linear likelihood function that supports data sets with flexible distributions that bear no parametric assumptions. Please visit the Amazon Forecast developer documentation for more detailed information.
This expanded hyperparameter support for DeepAR+ is now available in US East (N. Virginia, Ohio), US West (Oregon), Europe (Ireland),and Asia Pacific (Tokyo, Singapore, Seoul).