Abstract
Buildings account for a significant portion of global energy consumption, with HVAC and humidity-control systems representing the majority of their operational demand. Traditional rule-based strategies often fail to adapt to dynamic indoor-outdoor conditions, motivating the use of data-driven control methods. This study presents a multiagent reinforcement learning (MARL) framework for simultaneous temperature and humidity control in a singlezone building modeled in EnergyPlus. The proposed approach employs the distributed Importance Weighted ActorLearner Architecture (IMPALA) algorithm with centralized training and decentralized execution (CTDE), enabling two agents: temperature and humidity to learn coordinated policies directly from high-fidelity simulation feedback. The results demonstrate strong learning performance: both agents improved their per-step rewards substantially (temperature +18.9%, humidity +33.7%), indicating effective convergence and cooperative behavior. The learned controller maintained thermal comfort comparable to the rule-based baseline (mean occupied temperature difference ≈ 0.04 °C; occupied PMV ≈ 0.45) while achieving notable energy savings. Total annual HVAC energy consumption decreased by 8.9%, with the most significant improvement observed in humidification energy, which was reduced by 34.4%. Heating and cooling loads remained nearly unchanged, confirming that energy reductions were achieved without compromising comfort.