New Benchmark Boosts Generalizable Reinforcement Learning for HVAC Control.

Reinforcement learning (RL) has shown promise in control tasks, but its real-world application is often hindered by the brittleness of learned policies when faced with changes in dynamics, action spaces, or goals. Current benchmarks lack the diversity and complexity needed to thoroughly study critical RL challenges like transfer learning, multi-task learning, and meta-learning. To address this, a new benchmark called Building2Building (B2B) has been developed. It comprises a comprehensive suite of realistic Heating, Ventilation, and Air Conditioning (HVAC) control environments, built upon the advanced EnergyPlus building simulator. B2B integrates with the Gymnasium interface and includes a parametric building generator, allowing for the creation of diverse building configurations with varying observation and action spaces. This new suite defines benchmark tasks specifically designed to tackle open problems in RL, such as adapting to new goals, changing dynamics, shifts in action spaces, and cross-domain transfer. By offering a large-scale, diverse, and physically grounded testing ground with standardized evaluation, B2B is set to significantly advance research into generalization and transfer in continuous control, with the added benefit of enabling more energy-efficient HVAC systems.