Parametrizing low-power wireless (LPW) protocols is inher- ently complex. Protocol performance is affected by environ- mental conditions and hardware characteristics: therefore, to assess and compare the suitability of different parameter sets, one requires real-world testing (e.g., testbed experiments). At the same time, blindly testing all possible parameter con- figurations on a testbed (or performing an exploration giving sufficient statistical significance given the noisiness intro- duced by wireless experiments and real-world dynamics) is largely undesirable. We hence need solutions that allow one to identify the best parameter set for a given LPW protocol while minimizing the number of necessary real-world trials. In this context, we have recently designed different optimiza- tion frameworks capable of intelligently navigating a LPW protocol’s parameter space and identify the best parameter set with a limited number of real-world trials (e.g., APEX), also with the help of Large Language Models (LLMs). In our current solution, however, the identified best parameter set is specific to the characteristic of the environment in which the real-world trials are executed. As LPW protocol performance drastically varies across different environments, one would currently need to repeat the entire optimization process (which starts by randomly exploring parameter com- binations before converging towards optimality) to identify the most suitable parameter set for a new environment.
We aim to improve this process by investigating, for example, the transferability of knowledge collected in prior real-world trials to new environments
The idea is to extract information about the characteristics of the environment in which real-world trials were previously executed (such as the network topology, which is shaped by device placement and connectivity, as well as the traffic patterns and load), and provide it as an additional input to the optimization framework. By leveraging this existing knowledge, the framework could begin the search of suitable parameter sets with promising configurations instead of ran- dom exploration. This could speed up convergence and im- prove the overall efficiency of the optimization process. The goal of this project/thesis is hence to investigate methods for reusing available data sources, including traces, protocol arti- facts, and prior results, to support LLM-informed parameter optimization of LPW protocols, with the goal of accelerating convergence by leveraging relevant prior knowledge.