Using genetic programming and simulation to learn how to dynamically adapt the number of cards in reactive pull systems
Introduction
Pull production control systems aim at managing finished inventory and work-in-process (WIP) in order to satisfy customer demands in time while minimizing the related costs in the manufacturing process. They are generally based on the Just-In-Time (JIT) philosophy, whose objective is to deliver the right parts, at the right time, at the right place, and in the exact amount needed. The most well-known pull systems are probably Kanban (Lage Junior and Godinho Filho, 2010, Monden, 1981) and Constant WIP (ConWIP) (Prakash and Chin, 2014, Spearman et al., 1990), where production is allowed only upon the reception of authorization cards, used to control all the manufacturing process (Bollon et al., 2004, González-R et al., 2012). The former uses a loop of cards at each stage of the process and the latter is simpler, since it considers the whole process as a single-stage system in which each part is pushed through the system as soon as its production is allowed at the input of the system by a card. These two types of system are illustrated in Fig. 1. One important issue of such pull control systems is to determine the appropriate number of cards for each loop. This problem has been widely addressed using optimization approaches, which aim at finding those numbers, so as to maximize given performance objectives (see for example (Paris & Pierreval, 2001)). Unfortunately, the use of a fixed number of cards implies a stable production environment (Framinan & Pierreval, 2012), which is often not the case. Indeed, today the market changes and unpredictable fluctuations in demand occur. To face these major difficulties, numerous studies have proposed to dynamically adapt the number of cards, in order to render so-called token-based pull manufacturing systems (González-R et al., 2012) capable to adapt themselves to new operating conditions (Takahashi et al., 2004, Takahashi and Nakamura, 1999b).
Despite the widespread literature related to this problem, the development of adaptive control systems, whose purpose is to change dynamically the number of cards in each loop of the system, still represents a significant research challenge. Indeed, the stochastic nature of pull manufacturing systems and their complex dynamic behavior render the use of mathematical models to evaluate their performance not relevant if one wants to avoid restrictive assumptions. Moreover, determining when to add or remove cards in real time is a problem that is difficult to address using optimization since the system state evolves along time often in a non-predictable manner. In such cases, decisions are frequently not taken in advance, but in real time, often using heuristic strategies, which can be more or less complex, and more or less dependent on the state of the system (Coffman, 1976).
Artificial intelligence (AI), in particular machine learning, can be very useful to extract the necessary knowledge to make efficient decisions about adding or removing cards, and to make it accessible to decision makers, in view of their everyday use. Indeed, we are interested in learning rules of the following form:
Unfortunately, learning require the use of suited training sets, which turn out to be quite difficult to obtain for real-time decisions (Mouelhi & Pierreval, 2007). Providing examples or observations about the effect of a given decision, taken at time t, when the system is in a given state is generally extremely difficult since good or bad performances are induced by a sequence of coherent decisions taken at different instants of time. Moreover, the efficiency of decision sequences is generally difficult to measure on the very short term. As a consequence, one of the motivations of this research is to suggest a learning approach capable of generating decisions strategies, not requiring the use of such training sets, and that can be used for various pull control systems, without restrictive assumptions. In this respect, we propose to combine Genetic Programming (GP) and simulation, so that the knowledge needed to make efficient decisions is directly extracted from simulation runs. To the best of our knowledge, the joint use of these two techniques has not yet been studied in the literature to solve this kind of problem. The knowledge learnt can be implemented in the pull control system to determine when changes should be made and how many cards should be added or removed, or communicated to production managers who wish to improve their everyday practice.
The rest of this article is organized as follows. Section 2 analyzes the literature on adaptive pull control systems, and emphasis is put on articles concerned with learning techniques. Section 3 introduces our Simulation-based Genetic Programming approach. Section 4 provides an example adapted from the literature on adaptive ConWIP control, to which our approach is applied, and our results are discussed. Finally, our conclusions and research directions are drawn in Section 5.
Section snippets
Related research
Many articles have been devoted to the improvement of pull control systems and several states of the art published (Akturk and Erhun, 1999, Bollon et al., 2004, Di Mascolo et al., 1996, González-R et al., 2012, Lage Junior and Godinho Filho, 2010, Prakash and Chin, 2014). In the eighties, Monden (1981) underlined that Kanban systems should be used only in presence of small fluctuations. It is now well recognized that, when there are frequent and wide variations in supply and demand, then it may
General principles
We are interested here in suggesting a way to adapt the number of cards in each loop that composes a token-based pull system, in real time, depending on the current system state (e.g., number of parts in queue, number of cards in use, etc.), so as to satisfy given performance criteria. For doing so, a basic principle of our approach is that, rather than directly determining the number of cards to use at a given instant t, the decisions to be made concern a choice among the available candidate
Studied ConWIP reactive control system
The production system considered here is extracted from (Tardif & Maaseidvaag, 2001), as their model is one of the most well reported in the literature. It produces one single standard type of product, and assumes an infinite supply of raw parts. The process consists of one workstation with 10 machines in parallel where processing times are exponentially distributed with a mean of 6. As we have a dynamic system where the number of kanban cards is allowed to vary, it can use E extra cards in
Conclusion
How to react in the best way when customer demands change is a key question for most production managers. This quest for more adaptability is essential for companies, so as to maintain a high level of performance (Belisário & Pierreval, 2013), even if their manufacturing system is managed according to a pull control mechanism, based on the Just-In-Time philosophy. To cope with changes, determining when and how to modify the number of cards used in token-based pull control systems – which is
Acknowledgments
The authors would like to thank the μGP development team for the Free and Open Source Software μGP and, in particular, Dr. Alberto Tonda, research fellow at INRA, France and guest researcher at ISCPIF, France, for his kind help.
References (79)
- et al.
A parallel genetic programming based intelligent miner for discovery of censored production rules with fuzzy hierarchy
Expert Systems with Applications
(2010) - et al.
Prediction of the unified Parkinson’s disease rating scale assessment using a genetic programming system with geometric semantic genetic operators
Expert Systems with Applications
(2014) Knowledge-based simulation
Expert Systems with Applications
(1991)- et al.
Dynamic card controlling in a ConWIP system
International Journal of Production Economics
(2006) - et al.
An integrated MOGA approach to determine the Pareto-optimal kanban number and size for a JIT system
Expert Systems with Applications
(2011) Optimization and analysis aid via data-mining for simulated production systems
European Journal of Operational Research
(2006)Multi-robot path planning using co-evolutionary genetic programming
Expert Systems with Applications
(2012)- et al.
Optimization of simulated systems: OptQuest and alternatives
Simulation Modelling Practice and Theory
(2007) - et al.
Variations of the kanban system: Literature review and classification
International Journal of Production Economics
(2010) - et al.
Enhanced decision making mechanism of rule-based genetic network programming for creating stock trading signals
Expert Systems with Applications
(2013)
Self-adaptive systems: A survey of current approaches, research challenges and applications
Expert Systems with Applications
A rule induction approach for determining the number of kanbans in a just-in-time production system
Computers & Industrial Engineering
AI and machine learning techniques for managing complexity, changes and uncertainties in manufacturing
Engineering Applications of Artificial Intelligence
Expert system for selecting priority rules in flexible manufacturing systems
Expert Systems with Applications
Simplifying decision trees
International Journal of Man-Machine Studies
Hybrid expert systems: A survey of current approaches and applications
Expert Systems with Applications
Algorithm for the design of single-stage adaptive kanban system
Computers & Industrial Engineering
An application of the genetic programming technique to strategy development
Expert Systems with Applications
Reactive JIT ordering system for changes in the mean and variance of demand
International Journal of Production Economics
Decentralized reactive Kanban system
European Journal of Operational Research
An adaptive approach to controlling kanban systems
European Journal of Operational Research
Simulation optimization: A review on theory and applications
Acta Automatica Sinica
Reducing myopic behavior in FMS control: A semi-heterarchical simulation–optimization approach
Simulation Modelling Practice and Theory
An overview of design and operational issues of kanban systems
International Journal of Production Research
A multi-criteria decision making procedure based on neural networks for kanban allocation
Determining the parameters of dual-card kanban system: An integrated multicriteria and artificial neural network methodology
The International Journal of Advanced Manufacturing Technology
Genetic programming and redundancy
Unified framework for describing and comparing the dynamics of pull control policies
Annals of Operations Research
Improving induction of linear classification trees with genetic programming
Self-organization in distributed manufacturing control: State-of-the-art and future trends
Computer and job-shop scheduling theory
An analytical method for performance evaluation of kanban controlled production systems
Operations Research
Special issue on pull strategies in manufacturing systems and supply chains: Recent advances
Journal of Intelligent Manufacturing
Solving facility layout problems using genetic programming
Simplifying decision trees learned by genetic programming
Rapid modeling and discovery of priority dispatching rules: An autonomous learning approach
Journal of Scheduling
Token-based pull production control systems: An introductory overview
Journal of Intelligent Manufacturing
Cited by (32)
Novel Simulation Optimization Approach for Supply Chain Coordination and Management
2024, Procedia Computer ScienceA decision trees-based knowledge mining approach for controlling a complex production system
2020, Procedia ManufacturingAdaptive smart card-based pull control systems in context-aware manufacturing systems: Training a neural network through multi-objective simulation optimization
2019, Applied Soft Computing JournalCitation Excerpt :The available card is attached to the next product to be processed (see Fig. 6). In its adaptive version [16,26,44], the number of cards is systematically adapted to improve the system performance. The Kanban system uses a loop of cards at each stage of the production process.
Designing new heuristics for the capacitated lot sizing problem by genetic programming
2018, Computers and Operations ResearchCitation Excerpt :GP is still a young field of research and attracts a growing research community leading to several research works applying GP to a variety of problems ranging from natural sciences and engineering, to data classification (Gandomi et al., 2015) and finance (Chen, 2002). Applications in operations management are warehouse optimization (Karasek et al., 2014; Xie et al., 2014), production process control (Belisario and Pierreval, 2015), inventory control (Ghaddar et al., 2016; Kleinau and Thonemann, 2004) and particularly production scheduling where GP is used for seeking intelligent dispatching rules (Branke et al., 2016). However, to the best of our knowledge, a methodology to develop lot sizing rules by means of GP has not yet been published though the benefits of using GP are manifold:
A multi-attribute decision model for setting production planning parameters
2017, Journal of Manufacturing SystemsCitation Excerpt :The average total service level, backlogs and work in process for each simulation experiment was recorded, and the Nelson’s ranking and selection approach was applied to the results to compare the results and ascertain the best strategy. Belisário and Pierreval [7] proposes a simulation-based genetic programming approach to learn how to decide when modifying the number of cards in pull systems is worthwhile. The objective function corresponds to the expected value of inventory and backorder average costs.
Adaptive card-based production control policies
2017, Computers and Industrial Engineering