Smart heating, Agent, HomeAssistant, Node-RED, RaspberryPi, reinforcement learning

This paper presents a reinforcement learning-based heating control system for smart homes, utilizing Q-learning to optimize gas boiler operation. The system employs a single agent that evaluates multiple room states independently and determines appropriate heating actions. Each room state is represented by a vector composed of the target, current indoor, and outdoor temperatures. The agent uses an epsilon-greedy strategy to balance exploration and exploitation, with a five-minute delay to retrospectively assess the effect of each action. Rewards or penalties are assigned based on heating efficiency, temperature trends, and proximity to the target temperature. The system is implemented entirely in Node-RED without external machine learning libraries, ensuring transparency and low complexity. The experimental setup includes three thermostatic heads and one outdoor sensor, integrated via Home Assistant. A supervisory component aggregates the individual decisions and determines whether the gas boiler should be activated based on weighted priorities. The architecture is scalable with respect to the number of rooms, lightweight in terms of system requirements, and shows potential for integration into real-world smart home environments.

This paper presents a reinforcement learning-based heating control system for smart homes, utilizing Q-learning to optimize gas boiler operation. The system employs a single agent that evaluates multiple room states independently and determines appropriate heating actions. Each room state is represented by a vector composed of the target, current indoor, and outdoor temperatures. The agent uses an epsilon-greedy strategy to balance exploration and exploitation, with a five-minute delay to retrospectively assess the effect of each action. Rewards or penalties are assigned based on heating efficiency, temperature trends, and proximity to the target temperature. The system is implemented entirely in Node-RED without external machine learning libraries, ensuring transparency and low complexity. The experimental setup includes three thermostatic heads and one outdoor sensor, integrated via Home Assistant. A supervisory component aggregates the individual decisions and determines whether the gas boiler should be activated based on weighted priorities. The architecture is scalable with respect to the number of rooms, lightweight in terms of system requirements, and shows potential for integration into real-world smart home environments.