A Reinforcement Learning Approach to the Dynamic Job Scheduling Problem

Scheduling or day-ahead planning improves the efficiency of a process and often leads to other advantages such as energy savings and increased revenue. However, most real-world scheduling problems are very complicated and are usually affected by several external parameters. Hence, finding the best schedule given a set of jobs requires extensive calculations that increase exponentially with the number of jobs. Traditional schedulers are, at times, unable to address uncertainties in the system. This paper proposes a Reinforcement Learning approach for solving the Job Scheduling Problem in a dynamic environment with an aim to minimize the peak instantaneous electricity consumption. The training instance is randomly reset after a certain period and the solver uses online training to adapt to the new environment. Simulation results show that both the proposed approach and a Genetic Algorithm-based approach achieve the minimum peak power consumption possible, which is 58% less than on-demand dispatch. Also, for 82.2% of the simulations, our method finds a better schedule than its initialization.