Scalable Attention-based Reinforcement Learning Method for Multi-asset Control

Seyed Soroush Karimi Madahi, Giuseppe Gabriele, Bert Claessens, Chris Develder

Keywords: attention mechanism, electric vehicle, multi-agent control, reinforcement learning

TL;DR: This paper proposes a scalable centralized reinforcement learning approach with attention-based aggregation to jointly optimize a fleet of flexible assets.

Abstract: In this paper, we propose a scalable centralized reinforcement learning method to jointly optimize a fleet of flexible assets. The attention layer in our proposed architecture enables the agent to make decisions for each asset based on both its local and aggregated asset-specific information. As a proof-of-concept, we investigate the performance of the proposed method on an electric vehicle (EV) charging problem. The results show that the trained agent can effectively control multiple EVs to achieve a common objective (load flattening in our case).

Submission Number: 10