Discovering dispatching rules from data using imitation learning: A case study for the job-shop problem

Dispatching rules can be automatically generated from scheduling data. This paper will demonstrate that the key to learning an effective dispatching rule is through the careful construction of the training data, $$\{\mathbf {x}_i(k),y_i(k)\}_{k=1}^K\in {\mathscr {D}}$$ { x i ( k ) , y i ( k ) } k = 1 K ∈ D , where (i) features of partially constructed schedules $$\mathbf {x}_i$$ x i should necessarily reflect the induced data distribution $${\mathscr {D}}$$ D for when the rule is applied. This is achieved by updating the learned model in an active imitation learning fashion; (ii) $$y_i$$ y i is labelled optimally using a MIP solver; and (iii) data need to be balanced, as the set is unbalanced with respect to the dispatching step k. Using the guidelines set by our framework the design of custom dispatching rules, for a particular scheduling application, will become more effective. In the study presented three different distributions of the job-shop will be considered. The machine learning approach considered is based on preference learning, i.e. which dispatch (post-decision state) is preferable to another.