Rule-Based and Reinforcement Learning Approaches in Multi-Agent Systems for Indoor Microclimate Control

Abstract

This paper presents the design, implementation, and experimental validation of an edge-based agentic artificial intelligence system for adaptive smart home control. Unlike conventional rule-based automation, which relies on fixed threshold logic and predefined responses, the proposed approach models the home environment as a decision-making process in which an autonomous agent learns optimal control strategies through interaction with its surroundings. The system is deployed on a resource-constrained microcontroller platform and operates without reliance on cloud infrastructure, ensuring low latency, enhanced privacy, and reduced network dependency. Environmental parameters including temperature, humidity, illumination level, and weather-related signals are continuously monitored and encoded into a discrete state representation. Based on these states, the agent selects control actions for fan motor and LED lighting using a reinforcement learning strategy. The reward mechanism integrates comfort deviation and energy consumption into a unified objective function, enabling goal-directed behavior rather than reactive switching. The learning process gradually enables the agents to form stable control policies, resulting in convergence of environmental parameters toward their optimal values. Experimental evaluation conducted over a rule based and reinforcement learning approaches and extended condition based on microclimate parameters. The findings show that the agents progressively learn optimal control strategies, resulting in a compact clustering of environmental states around target values and demonstrating robust performance despite possible sensor noise and random action exploration. The proposed architecture illustrates how embedded reinforcement learning can transform traditional automation into a self-adaptive, memory-driven intelligent agent suitable for real-time smart home environments. The study confirms the feasibility of implementing agentic artificial intelligence on edge device ESP32 connected with sensors in porotype and highlights its potential for scalable, autonomous building control applications.