Abstract:

This research attempts to tackle the prevailing challenges in bandwidth estimation (BWE) for real-time communication systems, with a special emphasis on applying offline reinforcement learning to craft a more accurate neural network for bandwidth estimation than those built using traditional heuristics. The cultivated model, "CQLBWE", represents a data-driven approach to BWE, operating offline. The model exploits heuristic-based techniques of the past to formulate a proficient BWE policy. Furthermore, the successful usage of CQLBWE underscores the practicability of deploying offline reinforcement learning algorithms in the field of bandwidth estimation.