A Simulation-Based Learning Approach To Improve Responsiveness Of Industrial Scheduling

FACUNDO ARREDONDO; ERNESTO MARTÍNEZ

Valparaíso

Conferencia; 19th International Conference on Production Research.; 2007

The dynamic nature of shop-floor environments makes plant?s reactivity a crucial management issue. A challenge in the efficiency and effectiveness of production systems lies in improving responsiveness of their scheduling system. When an unexpected event happens, the production schedule may need to be updated in an attempt to diminish the impact of disruptions such as accepting/cancelling an order, raw material delay/shortage, variability of task durations or equipment failure. In this work, disruptive event handling is addressed as learning a ?rescheduling policy? using intensive simulations in the framework of reinforcement learning. To exemplify the simulation-based learning approach, the problem of optimal order insertion in an existing schedule is discussed. Insertion operators of different complexities are defined and the values of choosing them at different current schedule found by the incoming order are learnt using the Q-learning algorithm. To generalize Q-values across a continuum of schedule states and insertion operators a locally weighting regression technique is used. Results obtained highlight that simulation-based learning of order insertion heuristics is very appealing to increase real-time responsiveness of order negotiation and acceptance.