Robust seru production optimisation under uncertain worker processing times

The seru production system introduces a novel re-configurable production paradigm. Although random variations in worker processing times (WPTs) are inevitable due to fluctuations in worker efficiency and machine performance, few studies address uncertain factors in seru production problems (SPPs). Another challenge lies in the characterisation of nonlinear seru processing times (SPTs), leading existing exact algorithms to enumerate an exponential number of possible seru formations. To mitigate the uncertainty in WPTs, this paper proposes a decision-dependent robust optimisation model, which characterises both endogenous and exogenous uncertainties. The budget that regulates the conservatism of the uncertainty set is modelled as an affine function to decision variables. To tackle the robust SPP (RSPP) with infinite nonlinear constraints, we propose an exact equivalent reformulation and a dual-level robust Q-learning-based cooperative coevolution algorithm (DRQL-CCA). First, by analysing its structural properties, we propose the first equivalent mixed 0-1 second-order cone programming reformulation, which achieves up to one order of magnitude speedup compared with existing algorithms. Second, for large-scale problems, we design a robust Q-learning model to guide the search of the DRQL-CCA by incorporating the worst-case system performance as state and reward indicators. Moreover, the dual-level Q-learning approach provides a divide-and-conquer architecture by decomposing the decision-making process into sub-problem selection and optimisation stages. Compared with the deterministic model, the proposed robust model can significantly reduce the variance of the schedule. Numerical comparisons with state-of-the-art algorithms also verify the superiority of the proposed DRQL-CCA.