This repository contains the implementation of Meta-Reinforcement Learning (Meta-RL) for optimizing HVAC control in smart buildings using the Sinergym simulation environment. The project explores the Reptile algorithm, a Meta-RL technique, to improve adaptability across different building and weather conditions. The goal is to reduce energy consumption while maintaining thermal comfortability.

• Meta-RL Framework: Implements the Reptile algorithm for fast adaptation to new tasks.

• Task Distributions: Two approaches are explored:

  • Building-based: Training on different building types.
  • Weather-based: Training on varying weather conditions.

• Inner-Loop DQN: Uses Deep Q-Networks (DQN) for policy optimization within the Meta-RL framework.

• Sinergym Integration: Leverages the Sinergym environment, based on EnergyPlus, for realistic building simulations.

run the agent.py

to train using "weather" task distribution:

python agent.py reptile --train --mode weather

to train using "building" task distribution:

python agent.py reptile --train --mode building

to test the algorithm on the weather task distribution:

python agent.py reptile --mode weather

test on building distribution:

python agent.py reptile --mode building

You need to train before you can test.