Underwater communication receivers usually rely on the fine-tuning of numerous hyperparameters to perform optimally. This fine-tuning process is challenging and time-consuming, and must be carried out by domain experts. Using a receiver with a decision-feedback equalizer (DFE), it is shown that finding optimal hyperparameters can be formalized as a black-box optimization problem for which an objective function is minimized. This objective function is designed to address the degenerate states of the DFE, known as ``DFE hallucinations", where the DFE produces small apparent mean square errors yet very high bit error rate. We propose a method for automating the tuning of hyperparameters using a tree-structured parzen estimator (TPE) approach, an algorithm belonging to the large family of Bayesian optimization algorithms. Results obtained from synthetic and real channels demonstrate the efficiency of the method, allowing significant reduction of the packet error rate.