Introduction Efficient sensing in Internet of Things (IoT) networks often conflicts with limited battery capacity. Balancing sensing quality with energy use is a central challenge. This thesis investigates how a reinforcement learning approach can support this balance and enable more adaptive IoT systems. Research Question The main research question is: How can a tabular Multi-Objective Reinforcement Learning approach be used to effectively balance energy efficiency and sensing quality in dynamic IoT networks? Method The study follows the Design Science Research method and develops an artifact called TQ-IoT, a tabular Multi-Objective Reinforcement Learning (MORL) algorithm for IoT optimization. The research strategy is simulation, where an IoT environment is modeled with battery levels, event likelihood, and sensing outcomes. Data for evaluation are collected from repeated simulation runs and analyzed using descriptive analysis. Results The results show that TQ-IoT effectively balances energy consumption and sensing quality by adapting its actions depending on battery state and event likelihood. Energy-focused policies prioritize sleep actions and extend battery lifetime, while quality-focused policies increase sensing and transmission to improve detection performance at the cost of higher energy use. Intermediate preference settings give Pareto-efficient policies that achieve better sensing quality per unit of energy than baseline strategies. Discussion The findings confirm that a tabular MORL approach with Chebyshev scalarization is a viable framework for adaptive decision-making in energy-constrained IoT devices. By optimizing for both energy use and sensing quality, TQ-IoT enables flexible trade-offs that static or single-objective methods cannot achieve. However, the simplified simulation environment and tabular formulation limit generalizability. Future research should explore real-world deployments, adapt the model to support continuous state spaces, and provide mechanisms to handle dynamically changing preferences.
Zaki et al. (Thu,) studied this question.