Underwater optical wireless communication (UOWC) has emerged as a promising alternative for next-generation underwater communications, offering significantly higher data rates than acoustic links and overcoming the severe attenuation experienced b...
Underwater optical wireless communication (UOWC) has emerged as a promising alternative for next-generation underwater communications, offering significantly higher data rates than acoustic links and overcoming the severe attenuation experienced by radio-frequency signals in seawater. Among UOWC-based techniques, simultaneous lightwave information and power transfer (SLIPT) is particularly attractive because it can deliver both data and energy through a single light beam, opening up self-sustainable operation of underwater sensors. However, applying SLIPT in practical scenarios involving a remotely operated vehicle (ROV) and multiple energy-deprived underwater sensors (ED-USs) raises a challenging joint optimization problem. The time-switching ratio and power-splitting ratio of SLIPT, together with the mode-switching ratio of time-division duplex (TDD), must be jointly adapted while channel conditions vary significantly due to alignment errors, turbulence, and water-quality fluctuations. This joint optimization problem is highly non-convex, and model-based approaches typically require accurate channel state information and prohibitive computation. This thesis proposes an online algorithm that addresses these challenges by integrating meta-learning into a soft actor-critic (SAC) framework, referred to as Meta-SAC. The ROV is set as the learning agent, and each ED-US interaction is treated as a meta-task. The inner loop performs SAC updates that refine the policy for the current ED-US, while the outer loop adapts the entropy temperature parameter through meta-learning, providing favorable initialization for subsequent ED-USs. The reward function is designed with penalty terms and a safety margin to ensure that the SLIPT downlink rate threshold and laser-diode hardware constraints are robustly satisfied along with the maximization of the uplink sensing data rate.