The rapidly growing deployment of Edge AI devices performing high-demand tasks, such as real-time object detection, creates a critical challenge: balancing high performance (maintaining a target confidence) against the severe constraints of intermittent power supply from solar energy harvesting. This thesis addresses the necessity for a dynamic policy that can effectively manage this dual-objective trade-off over long operational horizons. The research establishes an empirical foundation via a parameter study conducted on Raspberry Pi hardware, quantifying the stochastic relationship between configuration parameters (model variant, resolution, frame rate) and actual power consumption/detection confidence, which revealed median shifts of up to 1.37 W in power consumption and up to 26 percentage points in detection confidence between different operational configurations. This data informed the construction of a custom Reinforcement Learning (RL) environment that utilizes Kernel Density Estimation (KDE) to model hardware stochastically and physics-based models for solar dynamics. To solve the dual-objective problem of maximizing performance while satisfying the long-term survival goals, a Proximal Policy Optimization (PPO) agent was trained within a Constrained Optimization framework. The agent's policy was rigorously evaluated over 24-hour and 48-hour cycles across six dynamic scenarios against static and random baselines. The results confirm that the PPO agent successfully learned an adaptive strategy: it consistently manages the trade-off better than non-learning baselines, strategically scaling its resource use based on real-time energy context. Quantitative analysis showed that the PPO agent survived up to 1.5 hours longer than the more power-hungry baselines while achieving at least 40 percentage points more SLA satisfaction than the least power-hungry static policy. This work provides a validated, data-driven approach for sustainable resource management in energy-constrained Edge AI systems.
Julia Oberauner (Mon,) studied this question.