Tuning an underwater communication link

We present machine learning algorithms to tune an underwater communication link. The link tuner is characterized by 3 features: a) It is data driven, rather than physics driven. Hence, it only needs bit error rate information as input and is independent of the modem implementation, b) The tuner balances exploration of the search space against exploitation of existing knowledge, and c) It optimizes for the average data rate, instead of searching for maximum possible data rate. We implement the link tuner on the UNET-II modem and present results from simulations, water tank tests and field trials. The results demonstrate a significant improvement in average data rate as compared to the average data rate attained without tuning.