Smart Scheduling of Home Appliances Using Happiness Aware Scalable Reinforcement Learning Agent

Debajyoti Dasgupta, Arijit Mondal, Partha Pratim Chakrabarti

Abstract: This study introduces IMPEARL (Incremental Model-based Predictor Network Agent with Reinforcement Learning), a novel deep reinforcement learning approach for optimizing household appliance scheduling. This method intelligently considers uncertainties, variable pricing models, and user satisfaction, efficiently scaling with increased tasks. The model uniquely represents appliance states and job statuses, employing a Deep Q-network with incremental training and feature-extractor networks for improved performance. By incorporating a recurrent connection, IMPEARL captures the time series aspect of the problem, leading to significant bill savings. Compared to traditional first-come, first-served methods and state-of-the-art models, IMPEARL achieves a 37.5% reduction in billing costs, a 66.5% increase in user satisfaction, a 29.08% improvement in minimizing billing costs, a 7.06% decrease in user dissatisfaction, and is 6.75% less sensitive to variability, demonstrating its effectiveness in dynamic, non-smart environments.