The operation of Water Distribution Networks (WDNs) requires methods (i.e. control algorithms) for operating actuators such as pumps, valves, and chlorine injections. Recently there has been some work in our group on applying reinforcement learning to these control tasks [1,2].
However, these works and others [3] have either focused on pumps or valves, not on controlling both simultaneously as needed in more realistic WDNs such as [4]. Depending on the types of valves and pumps, the action spaces can be continuous or discrete, possibly even binary. It is a challenge to combine these with appropriate RL algorithms.
The aim of this MA theses or projects is to investigate methods to apply RL to WDNs with Hybrid Action Spaces, possibly by using appropriate RL algorithms such as [5], but also by reducing or transforming the action space appropriately by respecting the physics equations governing these systems.
Keywords: Reinforcement Learning, Deep Learning, Critical Infrastructure, Water Distribution Networks
Literature