Model-based Data-driven Learning Algorithm for Tuning an Underwater Acoustic Link

Underwater acoustic channels are fast varying in both spatial and temporal domain and hence are characterized by non-stationary fading statistics. When the channel statistics change, a modulation scheme designed for a specific fading model will underperform which motivates the need for link tuning algorithms. In order to alleviate this problem, data-driven adaptive modulation techniques are studied previously. Since channel information is unknown, these algorithms solve the explore-exploit dilemma in order to take actions that result in maximizing the average data rate. Channel physics information is often ignored in the design of these algorithms. The information gained through channel physics such as delay spread, coherence time, doppler spread etc. of the channel plays an important role in narrowing down the search space of modulation scheme parameters. However, the channel physics by itself is not sufficient to find a good performing solution. Therefore, we develop a hybrid algorithm which utilizes both, the information gained from channel physics and techniques from data-driven algorithms to solve the explore-exploit dilemma. A simplified Orthogonal Frequency Division Multiplexing (OFDM) system is used to illustrate the concept and its parameters are tuned in an online fashion. In particular, an online learning algorithm is developed to track the goodness of the schemes and a multi-armed bandit like problem is solved for taking decisions sequentially in order to maximize the average data rate of an underwater acoustic (UWA) communication link.