Constrained Dimensionally Aware Genetic Programming for Evolving Interpretable Dispatching Rules in Dynamic Job Shop Scheduling

Abstract

This paper investigates the interpretability of the Genetic Programming (GP)-evolved dispatching rules for dynamic job shop scheduling problems. We incorporate the physical dimension of the features used in the terminal set of GP, and assume that the rules that aggregate the features with the same physical dimension are more interpretable. Based on this assumption, we define a new interpretability measure called dimension gap, and develop a Constrained Dimensionally Aware GP (C-DAGP) that optimises the effectiveness and interpretability simultaneously. In C-DAGP, the fitness is defined as a penalty function with a newly proposed penalty coefficient adaptation scheme. The experimental results show that the proposed C-DAGP can achieve better tradeoff between effectiveness and interpretability compared against the baseline GP and an existing DAGP.