NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·3h ago
AI Research
arXiv CS.AI·May 28·8 min read
OccuReward: LLM-Guided Occupant-Centric Reward Shaping for Demographic Equity in Grid-Interactive Buildings
TL;DR
Smart buildings that respond to electrical grid demands often optimize for energy efficiency rather than occupant comfort, creating inequitable outcomes across demographic groups. OccuReward uses large language models (LLMs) to automatically shape reward functions—the numerical signals that guide AI decision-making—to prioritize occupant well-being alongside grid responsiveness.
✦ Why It Matters
Engineers can use LLM-guided reward shaping to embed fairness into building automation systems without manual policy redesign.
Key Takeaways
- Use LLMs to automatically generate fairness-aware reward functions for building control systems.
- Incorporate occupant feedback and demographic constraints directly into reinforcement learning reward design.
- Evaluate fairness metrics across demographic groups when optimizing grid-interactive building performance.
- Balance grid responsiveness with occupant comfort equity through multi-objective reward shaping.
Full Summary
Grid-interactive buildings use reinforcement learning (RL)—a machine learning approach where AI agents learn by receiving numerical rewards for actions—to balance energy consumption with grid demands. However, standard reward functions typically prioritize grid benefits over occupant comfort, creating disparate impacts across demographic groups.
OccuReward leverages large language models (LLMs) to automatically generate and refine reward functions that explicitly account for occupant-centric fairness metrics. The method uses LLM reasoning to translate occupant preferences and demographic equity constraints into reward signals that guide RL agents.
By incorporating occupant feedback and fairness considerations into the reward design process, OccuReward enables buildings to maintain grid responsiveness while reducing comfort inequities. The framework was evaluated on thermal comfort scenarios, showing measurable improvements in fairness across demographic populations compared to standard grid-optimized baselines.