Multi-Agent Learning-Driven Cross-Layer Optimization for Adaptive Underwater Cognitive Acoustic-Optical Networks

The uniqueness of Underwater Cognitive Acoustic-Optical Networks (UCAONs) requires printing communication strategies to be resilient to dynamism in marine environments and yet be high-reliability, energy-efficient, and smart in resource usage. This paper presents a Multi-Agent Learning-Based Cross-layer Optimization Framework which is a cross-layer optimization scheme where physical, MAC and network layers are optimized by jointly making decisions at the underwater nodes. In contrast to the current available singlelayer or centralized optimization models, the proposed architecture considers each cognitive node as an independent learning agent that can sense local differences in channels such as a change in acoustic noise, optical scattering, attenuation by salinity as well as changes in mobility and adjust its actions to those of the nearby ones via a decentralized multi-agent reinforcement learning (MARL). An engine to cross-layer adaptation allows every agent to optimize its transmission mode choice, power control, spectrum access scheme and routing choices based on environmentally-driven reward messages and policy sharing among cooperating agents. The hybrid acousticoptical communication model is a dynamically controlled model that finds the best patterns of mode selection in the changing turbidity and bandwidth conditions by learning. Experiments with physics-based underwater channel models show that the offered framework has a great positive impact on throughput, communication delay, and energy consumption, as opposed to traditional DRL, heuristic, and fixed-layer optimization approaches. The findings demonstrate the possibility of introducing cross-layer MARL to facilitate autonomous, scalable, and resilient communication systems to support Internet of Underwater Things (IoUT) applications in the future, such as underwater robotics, underwater environmental sensing, and maritime security.