Application of stochastic real-valued reinforcement neural networks to batch production rescheduling

Abstract

This paper details the design and application of a hybrid neural network architecture for the rescheduling problem of batch manufacture. Design issues include the selection of an appropriate neural network paradigm, specification of the network architecture and support for multistep prediction. Application issues include decoupling the network dimension from that of the problem and the definition of suitable rescheduling operators. The ensuing hybrid network is tested against heuristics previously identified as typically representing estimates for best and worst case performance within a cross-section of batch rescheduling problems.