Multi-Objective Decision Optimization for IoT Sensor Scheduling

Sammanfattning

Introduction: Battery-powered Internet of Things (IoT) sensors must repeatedly decide whether to transmit data immediately or conserve energy for future operation. Frequent transmission keeps receiver-side information fresh but reduces device lifetime, while conservative transmission saves energy but increases information staleness. Fixed scheduling policies are limited because they commit the sensor to one operating point and cannot adapt to changes in battery level, channel quality, or sensor activity. This creates a need for a lightweight and configurable scheduling method that can balance information freshness and energy consumption. The problem underlying this thesis is that existing IoT scheduling approaches, whether static rules, heuristics, or single-objective learning methods, are locked to a fixed energy–freshness trade-off. Research Question: This thesis addresses the following research question: how can a tabular Multi-Objective Reinforcement Learning approach be designed and evaluated to balance Age of Information and energy consumption in energy-constrained IoT sensor scheduling? The work focuses on whether a learned policy can provide a better and more configurable trade-off than fixed-rule and adaptive heuristic baselines. Method: A tabularMulti-Objective Reinforcement Learning scheduler, named TQ-IoT, is developed for a single IoT sensor node. The scheduler maintains two independent Q-tables: one for energy cost and one for information freshness. These objectives are combined at decision time using Chebyshev scalarisation, allowing the operating preference to be changed without retraining. Information freshness is measured using Age of Information, which captures how old the most recently received update is at the receiver. The state space includes battery level, Age of Information, sensor volatility, and channel quality, and the available actions are sleep, wait, and transmit. The evaluation uses real FIT IoT-Lab M3 sensor, power, and radio traces from three hardware experiments. Trained policies are evaluated through trace replay and compared against seven fixed-rule baselines and one adaptive rule-based policy. Results: The results show that TQ-IoT produces a configurable trade-off between device lifetime and information freshness. Freshness-priority configurations reduce mean Age of Information compared with conservative fixed policies while maintaining longer lifetime than always transmitting. Energy-priority configurations approach the lifetime of conservative baselines while preserving fresher receiver information. Theweight-sweep evaluation shows that adjusting a single preference weight shifts the learned policy along the lifetime–freshness trade-off curve without retraining, which no fixed-rule baseline can replicate. A sensitivity analysis further shows that action-cost assumptions strongly affect apparent scheduler performance, demonstrating that unrealistic cost models can lead to misleading conclusions about both freshness and lifetime. Discussion: The findings indicate that a lightweight tabular multi-objective scheduler can provide an interpretable and runtime-configurable alternative to static IoT scheduling rules. The main limitation is that the available radio traces are dominated by poor-channel conditions, which prevents a full validation of the channel-quality state component. The study is also limited to a single-node setting and does not model multi-node contention. Future work should evaluate the framework under more diverse wireless conditions, extend it to multi-node deployments, and further investigate hardware-informed energy modelling. The thesis concludes that TQ-IoT can improve the practical energy–freshness trade-off in trace-replay evaluation while remaining simple enough for resource-constrained IoT scheduling scenarios.