Systems for heating, ventilation and air-conditioning (HVAC) of buildings are traditionally controlled by a rule-based approach. In order to reduce the energy consumption and the environmental impact of HVAC systems more advanced control methods such as reinforcement learning are promising. Reinforcement learning (RL) strategies offer a good alternative, as user feedback can be integrated more easily and presence can also be incorporated. Moreover, multi-agent RL approaches scale well and can be generalized. In this paper, we propose a multi-agent RL framework based on existing work that learns reducing on one hand energy consumption by optimizing HVAC control and on the other hand user feedback by occupants about uncomfortable room temperatures. Second, we show how to reduce training time required for proper RL-agent-training by using parameter sharing between the multiple agents and apply different pretraining techniques. Results show that our framework is capable of reducing the energy by around 6% when controlling a complete building or 8% for a single room zone. The occupants complaints are acceptable or even better compared to a rule-based baseline. Additionally, our performance analysis show that the training time can be drastically reduced by using parameter sharing.
