Low-Complexity Physics-Informed Reinforcement Learning Using Post-Decision States with Stochastic Sampling

Delay-sensitive Internet of Things (IoT) applications continue to grow in prevalence as new wireless technologies are adopted. Since these applications often operate in unknown dynamic environments, reinforcement learning (RL) has emerged as an effective method to learn optimal decision policies that improve their overall performance. However, typical data-driven RL techniques that have been adopted to solve these problems do not exploit available knowledge of system dynamics. Consequently, they must 'learn' some information about the system that may already be known to the system's designer. Post-decision state (PDS) learning, on the other hand, leverages known system information (i.e., it is 'physics-informed') to simplify the learning task and improve learning performance. However, this comes at the cost of increased computational complexity, and makes it impractical to implement on resource constrained devices. This work introduces stochastic PDS learning, a novel RL algorithm that combines traditional PDS learning with stochastic sampling to produce a physics-informed RL agent that can leverage known system information even with limited computational resources. Performance of stochastic PDS learning is compared against numerous traditional RL algorithms in the context of a delaysensitive energy-efficient scheduling problem simulated as an environment in Gymnasium.