Large Language Models (LLMs) have become increasingly influential in many fields due to their ability to generate sophisticated text and code. Trained on extensive text databases, these models can translate user requests into code snippets, design specific functions, and even create whole projects from scratch. They have numerous applications, including generating heuristic greedy algorithms for complex problems and formulating reward functions for robotics.
Due to their immense generative capabilities, leveraging LLMs for designing prompts that generate alternate algorithms is a logical step forward. However, there are significant challenges with this approach, most significantly the ability to produce high-quality algorithms directly for specified scenarios, which may be due to a lack of appropriate training data. LLMs are typically employed to produce collections of candidate algorithms with a range of designs, rather than a single optimized algorithm. Even then, it remains difficult for LLMs to adequately rank these algorithms to identify the best one.
Researchers from Microsoft Research, UT Austin, and Peking University have developed a system to tackle these challenges, known as LLM-ABR. This system uses LLMs to design adaptive bitrate (ABR) algorithms that are customized for varying network conditions. This unique application enables LLMs to design key elements such as conditions and neural network structures within a reinforcement learning framework.
The traditional method of designing ABR algorithms is notoriously complex and time-consuming, integrating heuristic, machine learning, and empirical testing approaches. The researchers used LLMs to bypass this complexity, generating a plethora of new designs by using input prompts and the source code of existing algorithms. An additional normalization check was added to the process to rectify the issue of the LLMs failing to perform normalization, leading to unnecessarily large inputs for neural networks. The remaining LLM-generated designs were subsequently evaluated, with the design which afforded the best video Quality of Experience (QoE) being selected.
Due to financial limitations, the paper’s network architecture design was limited to GPT-3.5, from which 3,000 network architectures were produced. A compilation check was then performed to filter out unworkable designs, with 760 designs successfully passing this stage and going on to be tested in a range of network scenarios.
The authors conclude that the LLM-ABR offers an innovative and effective solution, autonomously designing adaptive bitrate algorithms for a range of network conditions using LLMs’ generative powers. The applications of LLMs in producing ABR algorithms is of notable importance, as performance analysis of these code variants paves the way for the development of future ABR algorithms. The full paper is available for those interested in learning more about this exciting area of research.
