An Enhancement of Reinforcement Learning by Scheduling with Learning Effects

This paper present results, which reveal that approaches obtained for scheduling problems with learning effects can be successfully used to improve the quality of machine learning methods. It is illustrated by modelling some aspects of Q-learning agents as scheduling problems with the learning effect, and constructing sequencing and dispatching algorithms, which take into account the existence of learning. Their application to determine the sequence of tasks processed by Q-learning agents can visibly speed up their convergence to an optimal strategy. Furthermore, we show that a dispatch of tasks according to the longest processing time algorithm for parallel computing can be replaced by a more efficient procedure, if agents can learn. The numerical analysis reveals that our approach is efficient, robust and only marginally dependents on a learning model and an accurate approximation of task processing times.