Post by Anjukan Kathirgamanathan

Whilst 2020 certainly has been a crazy year, it's somehow flown by due to the similarly crazy amounts of time I have spent trying to figure out reinforcement learning from scratch. One of the motivations for this was the potential it has as a plug-and-play style controller for grid-interactive buildings to help with demand side management and decarbonisation of the electrical grid. My interest was piqued with the #citylearn challenge (https://lnkd.in/eSvjdug). Together with Kacper Twardowski from Computer Science at #UCD, we participated in this challenge and managed to finish in 2nd place. We utilised a centralised "Soft Actor Critic" that was able to outperform a manually tuned Rule-Based Controller and generalise over different buildings and climates. Our paper explaining our solution at #ACM #BuildSys20 #RLEM20 has just been accepted and I am happy to share it here (https://lnkd.in/e6esNQh). We hope to share some further insights and detail some of the challenges faced (of which there are many!) in the near future. #ESIPP #reinforcementlearning #demandsidemanagement #buildingenergyflexibility #smartgrid