# ReLBOT: A Transfer Learning Approach to Minimize Reinforcement Learning Risks in Smart Buildings

**Authors:** Mikhail Genkin, J. J. McArthru  
**arXiv:** [2305.00365](https://arxiv.org/abs/2305.00365)  
**Date:** 2023-04-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.00365v1/

## Key points

- Smart buildings aim to optimize energy consumption using AI algorithms
- Lack of historical data for training these algorithms during commissioning
- Online reinforcement learning (RL) algorithms show promise but carry risk of discomfort to residents
- Proposed technique called ReLBOT utilizes transfer learning and deep RL
- ReLBOT transfers knowledge from an existing optimized smart building to a newly commissioned one
- Aims to reduce adverse impact of RL agent's warm-up period
- Significant improvements demonstrated, including reduction in duration and prediction variance
- Research addresses challenges in initial exploration phase of RL algorithms in smart buildings
- ReLBOT minimizes risks while optimizing energy consumption levels through AI-driven algorithms.

## Summary

Smart buildings aim to optimize energy consumption by using artificial intelligent algorithms. However, when a smart building is commissioned, there is no historical data available to train these algorithms. Online Reinforcement Learning (RL) algorithms have shown promise in this regard, but their deployment carries a risk as the RL agent explores its action space and potentially causes discomfort to the building residents. To address this issue, the authors propose a new technique called ReLBOT. ReLBOT utilizes transfer learning in conjunction with deep RL to transfer knowledge from an existing optimized smart building to a newly commissioned one. By doing so, it aims to reduce the adverse impact of the reinforcement learning agent's warm-up period. The authors demonstrate significant improvements through ReLBOT, including up to 6.2 times reduction in duration and up to 132 times reduction in prediction variance for the reinforcement learning agent's warm-up period. This research by Mikhail Genkin and J.J. McArthru presents an innovative approach that addresses the challenges faced during the initial exploration phase of RL algorithms in smart buildings. By leveraging transfer learning and deep RL techniques, ReLBOT offers a solution that minimizes risks while optimizing energy consumption. This paper contributes valuable insights into improving the efficiency and effectiveness of smart buildings through AI-driven algorithms by reducing risks associated with online reinforcement learning agents' warm-up periods while also optimizing energy consumption levels.