A Reinforcement Learning Motivated Algorithm for Process Optimization

In process scheduling problems there are several processes and resources. Any process consists of several tasks, and there may be precedence constraints among them. In our paper we consider a special case, where the precedence constraints form short disjoint (directed) paths. This model occurs frequently in practice, but as far as we know it is considered very rarely in the literature. The goal is to find a good resource allocation (schedule) to minimize the makespan. The problem is known to be strongly NP-hard, and such hard problems are often solved by heuristic methods. We found only one paper which is closely related to our topic, this paper proposes the heuristic method HH. We propose a new heuristic called QLM which is inspired by reinforcement learning methods from the area of machine learning. As we did not find appropriate benchmark problems for the investigated model. We have created such inputs and we have made exhaustive comparisons, comparing the results of HH and QLM, and an exact solver using CPLEX. We note that a heuristic method can give a “near optimal” solution very fast while an exact solver provides the optimal solution, but it may need a huge amount of time to find it. In our computational evaluation we experienced that our heuristic is more effective than HH and finds the optimal solution in many cases and very fast.