Researchers at Microsoft AI Propose LLM-ABR: A Machine Learning System that Utilizes LLMs to Design Adaptive Bitrate (ABR) Algorithms

Large Language models (LLMs) have demonstrated exceptional capabilities in generating high-quality text and code. Trained on vast collections of text corpus, LLMs can generate code with the help of human instructions. These trained models are proficient in translating user requests into code snippets, crafting specific functions, and constructing entire projects from scratch. One recent application consists of developing heuristic greedy algorithms for NP-hard problems and creating reward functions for robotics use. Also, researchers use the power of LLMs to develop innovative networking algorithms.

Using LLMs to design prompts that directly generate alternate algorithms has great significance and common sense. However, it is very challenging for LLMs to directly generate high-quality algorithms for a given target scenario. One reason could be insufficient data to train LLMs for this particular task. Often, LLMs are used to generate a collection of candidate algorithms featuring diverse designs instead of generating an effective final algorithm. Still, it is challenging for LLMs to rank these algorithms and pick the best one. This paper resolves the problem by leveraging LLMs to generate candidate model designs and performing pre-checks to filter these candidates before training.

Researchers from Microsoft Research, UT Austin, and Peking University introduced LLM-ABR, the first system that utilizes the generative capabilities of LLMs to autonomously design adaptive bitrate (ABR) algorithms tailored for diverse network characteristics. It empowers LLMs to design key components such as states and neural network architectures by operating within a reinforcement learning framework. LLM-ABR is evaluated across different network settings, including broadband, satellite, 4G, and 5G, and outperforms default ABR algorithms consistently.

The traditional approach for designing ABR algorithms is complex and time-consuming because it involves multiple methods, including heuristic, machine learning-based, and empirical testing. To overcome this, researchers used input prompts and the source code of an existing algorithm in LLMs to generate many new designs. Codes produced by LLMs fail to perform normalization, leading to overly large inputs for neural networks. To solve this issue, an additional normalization check is added to ensure the correct scaling of inputs, the remaining LLM-generated designs are evaluated, and the one with the best video Quality of Experience (QoE) is selected.

In this paper, network architecture design is limited to GPT-3.5 due to budget constraints. 3,000 network architectures are produced by utilizing GPT-3.5, followed by a compilation check to filter out invalid designs, out of which 760 architectures pass the compilation check that is further evaluated in various network scenarios. The performance improvements from GPT-3.5 range from 1.4% to 50.0% across different network scenarios, and the largest gains are observed with Starlink traces due to overfitting issues in the default design. For 4G and 5G traces, although the overall improvements are modest (2.6% and 3.0%), the new network architecture consistently outperforms the baseline across all epochs.

In conclusion, the proposed model, LLM-ABR, is the first system that utilizes the generative capabilities of LLMs to autonomously design adaptive bitrate (ABR) algorithms tailored for diverse network environments. This paper includes the application of Large Language Models (LLMs) in the development of adaptive bitrate (ABR) algorithms tailored for diverse network environments. Further, an in-depth analysis is performed for code variants that exhibit superior performance across different network conditions and hold significant value for the future creation of ABR algorithms.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter. Join our Telegram Channel, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our 39k+ ML SubReddit