Using genetic programming and simulation to learn how to dynamically adapt the number of cards in reactive pull systems

Pull control systems are now widely used in many types of production systems. For those based on cards, determining their number is an important issue. When the system is submitted to changes in supply and demand, several researchers have demonstrated the benefits of changing this number dynamically. Defining when and how to do so is known as a difficult problem, especially when such modifications in customer demands are unpredictable and the system behavior is stochastic. This paper proposes a Simulation-based Genetic Programming approach to learn how to decide, i.e., to generate a decision logic that specifies under which circumstances it is worth modifying the number of cards. It aims at eliciting the underlying knowledge through a decision tree that uses the current system state as input and returns the suggested modifications of the number of cards as output. Contrarily to the few learning approaches presented in the literature, no training set is used, which represents a major advantage when real-time decisions have to be learnt. An adaptive ConWIP system, taken from the literature, is used to illustrate the relevance of our approach. The comparison made shows that it can yield better results, and generate the knowledge in an autonomous way. This knowledge is expressed under the form of a decision tree that can be understood and exploited by the decision maker, or by an automated on-line decision support system providing a self-adaptation component to the manufacturing system.