【Proposal】A Large Language Model-based Bandwidth Prediction Algorithm
Keywords: bandwidth prediction, time series, large language models
Abstract: In the streaming media industry, bandwidth prediction is vital for ensuring user experience and optimizing resources. In low-latency live streaming, it estimates network conditions in real-time, adjusts transmission strategies, and reduces stuttering and latency. For long videos, it helps adaptive algorithms intelligently select bitrates, balancing picture quality, smoothness, and buffering. In short videos, it determines video bitrate combinations for seamless HD playback, improving user retention. Bandwidth prediction also affects CDN distribution costs and efficiency by optimizing transcoding bitrates and scheduling. Accurate bandwidth prediction enhances decision-making, user experience, and technical architecture, crucial for competitiveness. Traditional algorithms struggle with long-term bandwidth variations due to user demands and network complexities. Recently, Transformer and Large Language Models (LLMs) from time-series prediction offer new solutions. Transformers excel in feature extraction and long-term dependency modeling, while LLMs adapt quickly using pre-training on large datasets. Applying these models to bandwidth prediction can greatly enhance accuracy, generalization, and real-time performance, addressing bandwidth prediction challenges more effectively.
Submission Number: 10
Loading