Multi-Tree Genetic Programming With Deep Contextual Bandits for Learning-Assisted Scheduling of Cluster Tools Integrating Stockers

This work proposes a dynamic scheduling problem of a single-armed cluster tool (SACT) by considering the wafer lots simultaneously stored in a stocker to be processed at it. In other words, the investigated system integrates a stocker and an SACT, which is named as an S-SACT system. Multiple wafer types (more than two) may be processed simultaneously by different processing chambers (PCs) in the SACT. Thus, PC cleaning is necessary when switching between wafer types. Stochastic lot arrivals necessitate dynamic decisions on lot sequencing in the stocker and their PC assignments, generating a partially-determined schedule that is updated when a new lot arrives. Then, an activity-driven simulation is used to obtain the final schedule. Generating high-quality partially-determined schedules is challenging. To address this, we propose a two-stage learning-assisted evolutionary framework combining Multi-tree Genetic Programming (MTGP) and Deep Contextual Bandit (DCB). In stage one, MTGP automatically generates scheduling rules that produce high-quality partially-determined schedules in dynamic environments offline, forming the action set for stage two. In stage two, DCB is pre-trained to select the most effective action with over 90% probability, then deployed to choose actions in real time while continuously updating its rule-selection agent. Computational experiments demonstrate that the proposed method consistently outperforms existing approaches across all tested cases.