A machine learning engineer is tuning a deep neural network using a SageMaker Automatic Model Tuning job. Each full training run takes several hours, and the team has a limited budget of 100 training jobs. Many hyperparameter combinations produce poor validation accuracy within the first few epochs but still consume the full training time. The engineer wants the tuning job to dynamically allocate more resources to promising configurations and stop unpromising ones early to maximize the number of configurations explored. Which tuning strategy should the engineer select?