This thesis uses a gradient-based evolutionary strategy (GES) to learn dispatching rules for the job shop scheduling problem. It has two main contributions: (i) analysing distribution-driven features (variance, skewness, kurtosis) and (ii) evaluating multiple-instance training on perturbed variants of one benchmark instance and across different benchmarks. For stochastic instances, this work shows that adding variance-based features yields distribution-dependent gains: under exponential processing times on small instances, the average improvement over the MWKR rule rises from about 4.3% to 4.6%, whereas additional skewness- and kurtosis-based features do not provide consistent benefits. Multiple-instance training on perturbed variants of a single instance improves robustness on disturbed test cases compared with single-instance training. Training across different instances shows that joint external training improves average robustness on perturbed instances by at most 2% relative to single-instance training, while base-instance performance deviates from the single-instance baseline by no more than 1.5%.     «

This thesis uses a gradient-based evolutionary strategy (GES) to learn dispatching rules for the job shop scheduling problem. It has two main contributions: (i) analysing distribution-driven features (variance, skewness, kurtosis) and (ii) evaluating multiple-instance training on perturbed variants of one benchmark instance and across different benchmarks. For stochastic instances, this work shows that adding variance-based features yields distribution-dependent gains: under exponential processi...     »