Presentazione in lingua inglese

Progettazione e Implementazione di un controllo con Meta-Reinforcement Learning



Meta-RL Control Architecture Introduction

Machine Learning - Overview

Supervised Learning: the ML algorithm learns an input- output mapping from a labelled dataset ?learn the model.
Unsupervised Learning: the ML algorithm is provided with unlabelled data and learns the underlying structure or distribution of the data ?learn the patterns in the data.
Reinforcement Learning: the ML algorithm learns which actions to take in order to maximize a numerical reward signal ?learn to control
A Reinforcement Learning (RL) agent aims to learn the optimal way to accomplish a task through repeated interactions with its environment, by evaluating the longtermvalue of its actions.

Approaches to Temperature Control
- Classical methods (PID)
- Advanced control (MPC)
- Data-Driven control (RL)

How to face uncertainty in the building dynamics?

Meta-Reinforcement Learning

Problem statement:
Design a temperature control system that rapidly adapts to uncertain building thermal dynamics - Meta-Reinforcement Learning (Meta-RL)

Meta Learning: provide machines with the skill to learn how to learn.

Characteristics:
- Learns over a distribution of environments
- Training offline
- fast adaptation to new / changing systems

Case Study Building 25

(in the PDF)