Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

A classification pattern for autonomous control methods in logistics

A classification pattern for autonomous control methods in logistics Logist. Res. (2010) 2:109–120 DOI 10.1007/s12159-010-0030-9 OR IGINAL PAPER A classification pattern for autonomous control methods in logistics • • • Katja Windt Till Becker Oliver Jeken Achim Gelessus Received: 9 February 2010 / Accepted: 15 May 2010 / Published online: 13 June 2010 Springer-Verlag 2010 Abstract Autonomous control in logistics enables single 1 Introduction logistics objects to control the production and transporta- tion process. This shift from central planning to decen- In the past few years, a change can be observed in the area tralized control in real-time offers many possibilities to of logistics. Technological developments and changing cope with highly dynamic and complex systems. The market conditions have resulted in rising complexity in algorithms that define the decision behavior of each production and consequently in logistics. These changes logistics object, autonomous control methods, play a key include an increasing number of product variants, faster role in the successful implementation of autonomous con- delivery times, and shorter product life cycles [7]. The trol in logistics systems. A transparent classification is complexity of nowadays logistics processes has significant needed in order to identify the basic elements these impact on the performance of logistics processes in terms methods consist of. This classification supports the evalu- of delivery time and delivery reliability [3, 5, 8, 18]. A new ation of autonomous control methods in terms of gain- approach to deal with complexity is to increase the level ing knowledge about which method characteristics are of autonomous control in logistics processes [10]. The responsible for a method’s performance. This paper defines Collaborative Research Center (CRC 637) ‘‘Autonomous what autonomous control methods are, works out their Cooperating Logistic Processes—A Paradigm Shift and its fundamental characteristics, presents multiple methods Limitations’’ in Bremen, Germany, aims at developing new developed so far, and compares these methods regarding methods in production planning and control (e.g. [12, 13, characteristics and performance. 17]) as well as in transportation control (e.g., [9, 15]) in order to overcome the obstacles created by today’s com- Keywords Autonomous control  Categorization  plexity and dynamics. Simulation studies have already Decentral  Logistics  Production planning and control shown that increasing the level of autonomous control improves the logistic performance [14]. At the same time, the availability of new information technologies such as any type of wireless communication (e.g. RFID), distrib- K. Windt  T. Becker (&)  O. Jeken uted computing, and computer miniaturization serves as an Global Production Logistics, School of Engineering and Science, enabler for autonomous control. The CRC focuses its work Jacobs University Bremen gGmbH, Campus Ring 1, on decentralized methods that have the ability to utilize 28759 Bremen, Germany given flexibility potentials in logistics processes [23], but e-mail: t.becker@jacobs-university.de also on gaining more knowledge on autonomously con- K. Windt trolled system behavior, e.g. identifying the limitations of e-mail: k.windt@jacobs-university.de autonomous control. A. Gelessus Several different autonomous control methods have Computational Laboratory for Analysis, Modeling, been developed and tested in simulation studies. These and Visualization, School of Engineering and Science, previously conducted studies have illustrated that auton- Jacobs University Bremen gGmbH, Campus Ring 1, omous control can realize a higher logistics target 28759 Bremen, Germany 123 110 Logist. Res. (2010) 2:109–120 achievement in comparison to conventional production as well as a scheme for characterization. From the above planning and control [9, 14, 25]. However, it remains mentioned description of autonomous control and its ele- unclear which basic characteristics make up autonomous ments follows this definition: control and what is their influence on the logistics system An autonomous control method is a generic algorithm performance. The goal of this work is to identify basic that describes how logistics objects render and exe- elements of autonomous control methods and their influ- cute decisions by their own. ence on performance so that future methods can be developed systematically by focusing on parameters that There are many different ways an autonomous control contribute to a higher achievement of logistics targets. method can operate. A simple method could allow each This paper provides an overview on the current under- semi-finished part to choose the next production step in a standing of autonomous control in logistics as well as, job-shop scenario by preferring that machine with the more specific, on the term autonomous control methods in lowest number of waiting items in front of a machine Sect. 2. By collecting the basic elements of these methods, [11]. Another example is a method that is inspired by a classification pattern is elaborated. In Sect. 3, a literature ants’ foraging behavior. It uses virtual pheromones, survey of currently available autonomous control methods emitted by the parts, which then indicate a preferred path is presented. The methods are classified in terms of the through the production [4]. The diversity of possible previously developed pattern, and a systematic similarity autonomous control methods leads to the questions of the analysis is conducted. Section 4 presents a comparison of methods’ individual performance and calls for a compar- the methods’ performance. The performance evaluation has ison of these methods. A previously developed Autono- been done using computer simulations. The conclusion mous Control Application Matrix is available to support brings together all insights and points out the next steps for the evaluation and comparison efforts [22]. A comparison research in this field. of simulation results is presented in Scholz-Reiter et al. [14], but only taking into consideration three different methods and aiming at showing the relation between 2 Autonomous control methods process complexity and level of autonomous control. Beside the evaluation of single methods, it is necessary to For getting a better understanding how autonomous control understand the structure of a method and to be able to works if applied in manufacturing and transportation identify similarities and differences in order to interpret environments or in computer simulations for scientific the comparison results. Additionally, the categorization purposes, it is necessary to consider the single elements presented in this paper can be used in the future to create a toolbox for autonomous control method development. that compose an autonomous controlled system. Autono- mous control itself is defined as follows: Upon availability of performance results for multiple methods, the performance can be connected to the method ‘‘Autonomous control in logistics systems is charac- characteristics and thus offers information which charac- terized by the ability of logistics objects to process teristics to consider when creating a method for a specific information, to render and to execute decisions on application purpose. their own.’’ [24] A recent approach for classifying autonomous control is The definition consists of different elements from dif- called the Catalogue of Criteria for Autonomous Control in ferent layers. On the one hand, there are the logistics sys- Logistics [2]. This catalog collects criteria from the deci- tem, the logistics objects, and the information, which are sion-making, information processing, and decision execu- structural elements that describe the environment in which tion. For each criterion, four possible properties are offered autonomous control takes place. Information processing for classifying the logistics system. The result of the and decision rendering and execution on the other hand are classification leads to a degree of autonomous control. This activities that characterize the way the flow of goods is degree describes—taking into consideration the thirteen controlled in such a system. Thus, two different layers can different criteria—how much a system is autonomously be observed in a logistics system: the logistics processes as controlled and thus makes different logistics systems well as the control methods that operate on the logistics comparable in terms of the usage of autonomous control. processes. In this paper, the focus is set on control methods For the purpose presented in this paper, namely making the that enable logistics objects to take their own decision in building blocks of autonomous control methods visible, contrast to a scheduled and centralized production planning this approach is not feasible as it only offers the degree of and control strategy. autonomous control as a single figure but is not able to A prerequisite for the survey presented in Sect. 3 is compare multiple methods regarding their basic elements having a definition of the term autonomous control method of their construction. 123 Logist. Res. (2010) 2:109–120 111 Coming from the methods that have been analyzed, autonomous control methods are kept simple and have a characterization of autonomous control methods can be very short information horizon, which corresponds to done by using seven dimensions that have been defined: having 1 planning step, but some algorithms additionally temporal data, planning steps, artificial values, communi- take into consideration what can happen two, three, or cation type, data scope, actor, and data storage. Each more steps after the pending decision. Artificial values dimension is separated in reasonable sections, e.g. planning are figures that are not extracted directly from the envi- steps can be described in discrete steps as 1, 2, 3, or more ronment but are generated by the algorithm itself. These planning steps. For categorization purposes, the dimension artificial values can be pheromones as presented in the values have been arranged in up to four groups. An over- above mentioned ant approach, but also virtual money view of dimensions, values, and descriptions can be found used in auction-theoretic approaches. The Communication in Table 1. type describes how communication is conducted by the The dimension Temporal data indicates whether the algorithm. It distinguishes between communication method uses data from the past, future, or both. Past data among the moving logistics objects (parts), among the are defined as any kind of figure that can be extracted fixed objects (machines), among all objects, or commu- from the environment, e.g. buffer inventory level of a nication with a central control entity. The communication machine, moving average of processing time at a itself can range from a simple data request (e.g. a part machine, etc. Future data can be planned data or esti- demands the current buffer inventory level from a mated values for the above mentioned figures. Caution machine) to extensive negotiations between agents (e.g. should be exercised at this point as estimated values are parts bidding in an auction for resources). The Data often determined by calculating averages from past data, scope provides information about the number of variables and thus, these kind of values have to be classified as used for decision making. The dimension Actor points past data. The number of Planning steps points out how out whether the parts, the machines, or a central entity deep the algorithm searches in a decision tree. Usually act as decision maker in the algorithm. An issue regarding this dimension is the fact that from the decision point of view, it is not important who takes the decision. Only the input parameters and the target system define Table 1 Autonomous control categorization dimensions the outcome of the decision. This dimension is never- Dimension Values Description theless part of the categorization, as it plays an important role when implementing an autonomous control method Temporal data Past Indicates whether the method uses data from the past, future in a real logistics environment. The hardware layout of Future (planned), or both an autonomous control solution is strongly influenced by Hybrid this characteristic. Finally, Data storage describes where Planning steps 1 Number of future steps (e.g. the figures are stored that form the basis for decision machines) the method considers making. Again, this dimension is closely related to the hardware implementation. More Artificial values No Usage of artificial values for decision making, e.g. virtual Static pheromones 3 Autonomous control methods survey Dynamic Communication Part-machine Communication and data exchange In this section, autonomous control methods presented in type between logistics objects or a Part–part central entity and logistics the literature are described, characterized in accordance Machine– objects with the above defined dimensions, and compared by per- machine forming a systematic similarity analysis. Central Data scope Low (1–2) Number of variables used for 3.1 Methods description decision making Medium (3–5) High ([5) Various methods for autonomous control in logistics are Actor Part Logistics object that actively available today. Some of them have been developed in decides Machine the course of the research of the CRC 637, while others Central were invented without explicitly naming them autono- Data storage Part Location of data storage mous control methods. The descriptions of each single Machine autonomous control method identified can be found in Central Table 2. 123 112 Logist. Res. (2010) 2:109–120 Table 2 Autonomous control methods survey including method name (short name in parentheses), key idea, basic algorithm description, and remarks Method Ant algorithm (Ant) Cunning ant system (C-Ant) Source Ant colony control for autonomous decentralized shop floor The cunning ant system by Tsutsui and Pelikan [16 ] routing by Cicirello and Smith [4] Key ideas Ants carry products Two different types of ants exist Different jobs have different types of pheromones c-Ants, cunning Ants base most of their path on previous paths Machines have pheromone concentrations d-Ants are conventional ants, who donate their path to the c-Ants Ants choose machines based on pheromone concentration Both locate pheromones at spots on their path Pheromones expire over time different pheromones will be overwritten Optional: ants sometimes choose machines randomly pheromones evaporate after some time Algorithm 1. Initially choose a random machine 1. Let in some d-ants and some c-ants 2. Avoid machines with different pheromones 2. Let the d-ants choose the paths with the higher pheromone concentration with a higher likelihood 3. Go to machines with same pheromones 3. Let already deposited pheromones expire at a constant rate 4. Let pheromones expire after some time 4. Let c-ants retrieve the paths of the d-ants and use them partially for their own paths 5. Increase pheromone concentration if ant has visited the machine 6. Overwrite old pheromones if the pheromone type of the ant is different 7. Update pheromone concentrations once ant is through 8. Optional: choose a random machine Remarks Long initiation time; Algorithm avoids stagnations and is thus more flexible than the Not flexible if job influx changes ordinary ant algorithm Method Pheromone approach (Ph) Bee foraging (Bee) Source Autonomous control of production networks using Autonomous control of a shopfloor based on bee’s foraging a pheromone approach by Armbruster et al. [1] behaviour by Scholz-Reiter et al. [13] Key ideas Average throughput time is used as a pheromone Bees indicate good food sources via dancing To model evaporation, only the last n throughout Food sources are measured by quality and quantity times are considered Number of relevant parts for the throughput time More bees go to the better sources is the rate of evaporation Algorithm 1. Calculate average throughput times for the last n parts 1. Choose the best food place 2. Go for the machines with the lowest average 2. Different food sources (machines) exist throughput times a. Advertise (dance) and send information aa. Number of recruited bees depends on the number of dances ab. The quality of the source depends on the length of the dance b. Just use but don’t recruit c. Abandon machine and join pool of unemployed bees Remarks Slow adjustment to change Long initiation time Not flexible if job influx changes Method Simple rule based 1 (SRB 1) Simple rule based 2 (SRB 2) Source Autonomous control of a shop floor based The Influence of Production Networks’ Complexity on the on bee’s foraging behaviour by Performance of Autonomous Control Methods Scholz-Reiter et al. [13] by Scholz-Reiter et al. [11] Key ideas Compares estimated waiting time at buffers Compares estimated waiting time at buffers Uses future events Uses data from past events Algorithm 1. Parts are rated in estimated processing time 1. When a part leaves a machine it sends information about the processing times 123 Logist. Res. (2010) 2:109–120 113 Table 2 continued Method Simple rule based 1 (SRB 1) Simple rule based 2 (SRB 2) 2. Current buffer levels are calculated as a sum 2. This information is used by the following parts of the estimated processing time to decide where to go next 3. Choose the machine with the 3. Parts choose the machine which provides the lowest lowest processing time buffer mean duration of waiting and processing for parts of the same type Remarks Useful with high number of machines Useful with high number of products; Changes more slowly compared to the previous method Method Queue length estimator (QLE) Due date (DD) Source Modeling and analysis of production logistics processes Modeling and analysis of production logistics processes based on biologically inspired strategies by J. Bendul based on biologically inspired strategies by J. Bendul (Master’s thesis, University of Bremen, Germany) (Master’s thesis, University of Bremen, Germany) Key ideas Computes and estimates buffer states Uses the queue length estimator (QLE) Part decides autonomously based on various factors Orders parts by earliest due date Algorithm 1. All buffer states of machines that can perform 1. After a part leaves a machine it chooses its next the next step are computed destination based on the QLE method 2. The part decides whether to switch to a different 2. Within the queue of parts to be processed the part production line based on processing times or setup with the most urgent due date is chosen times, using local information to be processed next 3. Parts compare their own estimated time with the estimated time of the parts in the buffers and takes the machine with the minimal time Remarks Similar to the simple rule methods above Similar to the simple rule methods above Method One logistics target per rule (OLTPR) Market based control (market) Source Developed in this research group Developed in this research group according to Vollmer [20] Key ideas Implement various rules at the machines and parts, where Virtual currency is introduced each rule tried to achieve a specific logistics target Can be easily extended with new rules to further Parts carry a shopping list of work that needs improve outcome to be done on them Each job needed for a part has a budget associated Distance traveled to the machine has a price Parts auction for access to the machine Shopping List can be altered during the production process Algorithm 1. Utilization: each machine send a stronger attraction 1. Parts with shopping list and budgets enter the signal as its buffer becomes less full production process 2. On time delivery: parts are prioritized by their due date 2. Parts bid on the machines on their shopping list 3. On time delivery: parts prefer machines with 3. Highest bidder gets access to the machine short throughput time 4. The parts bid according to the minimal price of the machine and the distance cost 5. Machines grant access for the parts, if they are the highest bidder Remarks The various rules have to be weighted appropriately It is not clear how to choose the budgets to achieve good performance Price levels on the machines are important for production activity Development of price levels might be usable to investigate overall state of the production process, thus provide macro data Production is more dynamic as shopping lists can be altered during the process 123 114 Logist. Res. (2010) 2:109–120 Table 2 continued Method Holonic manufacturing (Holonic) Bionic manufacturing system (Bio) Source According to van Brussel [19]; A market approach to holonic Reinforcement learning approaches to Biological Manufacturing manufacturing by Markus et al. [6] Systems by Ueda [17] Key ideas Two agents bargain over the next item to be processed Attraction fields dependent on the type of job exist Agents are machines and management Fields attract specific jobs Management and machines bargain for the jobs to do Jobs have DNA like information about what work needs to be done on them Management punishes machines for delays Machine have operating knowledge that evolves Machines bid to get jobs from management Algorithm 1. After a part leaves a machine it chooses its next 1. Machines(Robots) are attracted by fields destination based on the QLE method 2. Within the queue of parts to be processed the part 2. Parts send out fields, depending on the production with the most urgent due date is chosen process they require to be processed next Remarks Punishment have to be chosen carefully No specific algorithm provided Existence of social dilemma; machine decision Based on manufacturing processes that require robots might cause overall loss, but gain for machine and on the spot machines Requires a central authority Method Link-state internet routing protocol (LSIRP) DLRP (DLRP) Source Developed in this research group according Autonomous control by Means of Distributed Routing to Wenning et al. [21] by Wenning et al. [21] Key ideas Based on a link-state routing protocol Parts request a route from machines Each machine has a map of the entire facility Machines communicate best routes to a destination Parts can be sorted according to any rule Algorithm 1. A map of the facility and the connections between 1. Each machine is a knowledge broker machines is built/provided 2. Shortest paths are computed based on various chosen 2. Before a part decides on a best route it ask the criteria, generally using Dijkstra’s algorithm current machine about possible ways to reach the destination 3. As the situation changes (breakdowns, buffer states, 3. Each machine includes relevant information from new machines) only the changes are propagated its knowledge base and forwards it to its successors among the machines 4. To make sure that insignificant changes are not 4. The successors do the same and forward this information propagated there must be a lower threshold along the production chain 5. The request is propagated through the network until the destination (drain) is reached 6. Then the last broker (the drain or the last machine) sends a reply directly to the part with all the collected information 7. After receiving one or multiple probable paths the part decides on the better way to take Remarks The machines can self-organize if information about A timeout may be included to reduce the amount the part production cycle is included of waiting for possible paths The threshold may need to be scaled by the Can be computationally expensive local buffer level 3.2 Systematic comparison of autonomous control the similarity between two methods. For two autonomous k l kl methods control methods M and M , the similarity is labeled S . Similarities can be defined for each categorization dimen- A systematic comparison of two autonomous control sion, and the similarity for the i-th categorization dimen- kl kl methods can be achieved by a mathematical definition of sion is labeled S . The total similarity S between two 123 Logist. Res. (2010) 2:109–120 115 k l autonomous control methods M and M is defined as the It should be mentioned that the temporal data M and arithmetic mean of the categorization-based similarities data scope M can also be represented as nonexclusive data kl kl 1 kl pairs (past/future) and (low/high), respectively, which leads S : S ¼ S . The categorization-based similari- i i n i¼1 kl kl kl kl to a similar definition of S and S as for S . Thus, for 5 ties S are defined for each categorization dimension 1 5 4 out of 7 categorization dimensions, an identical similarity individually. kl definition is used. The similarity S for temporal data is given by the The classification of the autonomous control methods following similarity matrix: with respect to the newly defined categorization dimen- Past Hybrid Future sions (see Table 1) is shown in Table 3. So far, the Past 1:00:50:0 ð1Þ Bionic Manufacturing System method (Bio) is conceptual Hybrid 1:00:5 research only, and a classification is not possible at the Future 1:0 moment. For the other 13 autonomous control methods, a According to this definition, two methods are completely classification according to the categorization dimensions similar with respect to the temporal data if they use data is possible. These data are used to calculate the simi- kl k l from the same temporal range. Distinctions about the data larities S for all pairs (M , M ). The full similarity number, origin, and quality are not made in the present matrix is shown in Table 4. For symmetry reasons, only analysis. the upper triangle is shown, and the diagonal elements, kl The similarity S for planning steps is defined by a 4 which are always 100%, are omitted. The SRB1 method 9 4 similarity matrix: and the QLE method have a similarity of 100%. This does not mean that both methods are identical. The ð1Þð2Þð3Þð [ 3Þ classification of the autonomous control methods is based ð1Þ 1:00:50:333 0:0 on a rather rough scheme, and small differences in any ð2Þ 1:00:50:0 ð2Þ of the categorization dimensions are not observable in ð3Þ 1:00:333 the analysis. ð [ 3Þ 1:0 Data from Table 4 are used to identify clusters of similar The present analysis does not make a distinction for methods. A method cluster is defined as a set of autono- methods using more than 3 planning steps. kl mous control methods with all similarities S larger than kl The similarity S for artificial values is defined by the the threshold 71.43%, which corresponds to an identical following similarity matrix: classification for 5 out of 7 categorization dimensions. Four No Static Dynamic autonomous control method clusters could be identified No 1:00:00:0 (see Fig. 1 for a graphical representation). The categori- ð3Þ Static 1:00:5 zation values for the four clusters are shown in Table 5. Dynamic 1:0 The clusters 1, 2, and 3 overlap, and 5 methods belong to more than one cluster. This indicates that all 8 methods Autonomous control methods using artificial values (Ant, Bee, DD, OLTPR, Ph, QLE, SRB1, and SRB2) of the have no similarity to methods using no artificial values. clusters 1, 2, and 3 are similar, which is confirmed by Autonomous control methods using static artificial values inspection of the categorization values (Table 5). Common are defined as 50% similar to methods using dynamic features for the autonomous control methods of the clusters artificial values. 1, 2, and 3 are For the communication type, the four values are non- exclusive, and more than one communication type can be – Planning Step is 1 realized in an autonomous control method. The similarity – Part–Machine communication is used kl S for the communication type is defined as the arithmetic 4 – Data Scope is either low or medium mean of the similarities for all communication types: – Decisions (Actor) are made by parts 4 – Data Storage is on the machine k l 1 1 : M ¼ M kl kl kl 4j 4j S ¼ S with S ¼ ð4Þ 4 4j 4j k l The Pheromone Approach fulfills the similarity criterion to 0 : M 6¼ M 4j 4j j¼1 all methods of the clusters 1, 2, and 3, and the methods Ant, kl The similarity S for data scope is defined by the same Bee, DD, OLTPR, QLE, SRB1, and SRB2 can be regarded kl similarity matrix as S with the labels replaced by low, as modifications of the Pheromone Approach although medium, and high. most methods have a completely different derivation. The kl kl The similarities S and S for actor and data storage, methods LSIRP and DLRP create cluster 4, which deviates 6 7 kl respectively, follow the as definition as for S , but the from cluster 1, 2, and 3 in the following categorization arithmetic mean is created from three contributions. values 123 116 Logist. Res. (2010) 2:109–120 – Planning Step is larger than 3 – Machine–Machine communication is used – Data Scope is high – Decisions (Actor) are made by machines The methods C-Ant, Market, and Holonic cannot be assigned to any of the method clusters, and each method is a singular development so far. The cluster analysis shows that many of the autono- mous control methods developed so far bear much resemblance to each other despite their derivations from different origins. So far, rarely used categorization values could be the starting point for the development of new autonomous control methods that differ significantly from the existing ones, and the following features should be considered: – Planning Step is larger than 1 – Artificial Values are used – Communication is not restricted to Part–Machine – Data Scope is high – Decisions (Actor) is not restricted to parts – Storage is not restricted to machine 4 Simulation studies After classifying and analyzing the different methods, a comparison regarding their performance will be presented in the following. To compare different autonomous control methods, a generic and common production scenario is used. The shop floor consists of m parallel production lines where each production line comprises n machines M ij (i = 1, …, m, and j = 1, …, n). Each machine has an input buffer B with a maximum level of 40 items. K is the ij number of different products that can be produced by the production system. At the source, the raw materials for each product type enter the system. The input frequencies for the raw materials are expressed by three temporally shifted sinus- shaped functions. This input model assumes a continu- ously varying number of incoming orders and reflects the dynamics within the production system. It is assumed that the different products have different operation times on the machines. The priority rule applied at the machines is first-come-first-served (FCFS) if no other rule is applied by the respective autonomous control. As the scenario is related to the classical job-shop factory layout, all production lines are connected at every stage. Every line is able to process every kind of product, and the parts can switch to a different line at each production stage. The product’s choice of the next processing machine is determined by the applied autonomous con- trol method. Table 3 Autonomous control methods classification Ant C-Ant Ph Bee SRB 1 SRB 2 QLE Temporal data Past Past Past Past Future Past Future Planning steps 1 More 1 1 1 1 1 Artificial values Dynamic Dynamic No Static No No No Communication type Part-machine Part-part Part-machine Part-part, part-machine Part-machine Part-machine Part-machine Data scope Low High Medium Medium Low Low Low Actor Part Part Part Part Part Part Part Data storage Machine Part, machine Machine Machine Machine Machine Machine DD OLTPR Market Holonic Bio LSIRP DLRP Temporal data Future Hybrid Past Future — Past Past Planning steps 1 1 1 3 — More More Artificial values No No Dynamic Dynamic — No No Communication type Part-machine, part-part Part-machine, part-part Part-machine, central Central, machine-machine — Machine-machine, central Machine-machine, part-machine Data scope Medium Medium High High — High High Actor part Part Part, machine Part, machine, central Machine, central — Machine Part, machine Data storage Part, machine Part, machine Part, machine, central Machine, central — Machine, central Machine Logist. Res. (2010) 2:109–120 117 Table 4 Autonomous control methods similarity in % C-Ant Ph Bee SRB 1 SRB 2 QLE DD OLTPR Market Holonic LSIRP DLRP 60 79 82 71 86 71 56 58 63 36 32 49 Ant 52 63 31 45 31 46 49 61 45 56 65 C-Ant 82 79 93 79 77 80 56 29 54 70 Ph 61 75 61 67 69 60 32 36 52 Bee 86 100 85 73 35 36 32 49 SRB 1 86 70 73 49 21 46 63 SRB 2 85 73 35 36 32 49 QLE 88 43 35 31 48 DD 55 32 43 60 OLTPR 63 50 50 Market 54 37 Holonic 83 LSIRP Cluster 2 output quantity during the simulation run is counted. Fig- Cluster 4 ure 2 shows the results. Except for the Ant and SRB2 algorithm, the investigated methods show a linear increase LSIRP Ant SRB 2 SRB 1 in the output quantity for both cases with and without QLE DLRP Bee machine failures. This shows the general performance but Ph OLTPR does not take into consideration the achievement of the different logistics targets. The logistics targets are short DD lead times, high due date reliability, high utilization, and Cluster 1 Cluster 3 low inventory [7]. In the simulation, the lead time is measured as the mean throughput time of the parts in the manufacturing system. The due date reliability is measured Holonic C-Ant as the standard deviation of the throughput time, because it Market is assumed that low standard deviation indicates high predictability of lead time and results in high due date Fig. 1 Autonomous control method clusters performance. The utilization is measured as percentage of time the machines are busy in relation to the total time To compare the different methods, simulations were capacity. The inventory level is measured as the average conducted for production scenarios from m = 3to m = 9 number of parts in the buffers in front of the machines. with m = n and K = 3 (products A, B, and C). The Figure 3 shows the mean throughput time for the dif- operation time was determined in relation to the simulation ferent methods. Without machine failures, the distribution time. Product A had an operation time of 0.1%, product B appears evenly among the methods, having QLE, Holonic, of 0.5%, and product C of 1% of the total simulation time. LSIRP, and DLRP performing in a similar pattern, whereas To analyze the performance of different control methods, Ant and SRB2 perform significantly worse (see Fig. 3a). In all methods that have been described in detail by their case of machine failures, SRB2 still performs worse than publication sources were selected. As the autonomous the other methods, but the difference compared to the other control methods are designed to cope with complex and methods is smaller as displayed in Fig. 3b. The standard dynamic manufacturing systems, the comparison is carried deviation of the throughput time shows again the distinct out for scenarios with and without machine breakdowns. difference between the SRB2 and Ant method in relation to The breakdowns of the machines represent an additional the rest of the investigated methods (see Fig. 4). However, kind of dynamics and unpredictability of the system beside again, the difference is smaller in the case of machine the input fluctuation. The breakdowns occurred randomly failures than that without machine failures. The two with a probability of 75% and a repair time of 1 min. machine utilization graphs (see Fig. 5) show a similar In order to determine the performance of the different pattern of performance where the drop in the performance autonomous control methods, several key indicators are in Fig. 5b is caused by the machine failure. The previously measured. All results presented here are average values mentioned well-performing group of methods achieves to over five independent simulation runs. First, the overall keep the utilization on a high level, whereas with an 123 118 Logist. Res. (2010) 2:109–120 Table 5 Common characteristics of the method clusters identified Methods Cluster 1 Cluster 2 Cluster 3 Cluster 4 Ant, Ph, Bee, SRB 2 Ph, SRB 1, SRB 2, QLE, OLTPR Ph, SRB 1, QLE, DD, OLTPR LSIRP, DLRP Temporal data Past Different values Different values Past Planning steps 1 1 1 [3 Artificial values Different values No No No Communication type Always including Always including Always including Always including part-machine part-machine part-machine machine-machine Data scope Low/medium Low/medium Low/medium High Actor Part Always including part Always including part Always including machine Data storage Machine Always including machine Always including machine Always including machine Fig. 2 Total output of Comparison of AC Methods without failure Comparison of AC Methods with failure (a) (b) simulation run. a Without 3000 3000 machine failure, b with machine 2500 2500 failure 1500 1500 1000 1000 500 500 0 0 2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16 # machines n X n # machines n X n SRB2 Holonic DLRP SRB2 Holonic DLRP QLE LSIRP Ant QLE LSIRP Ant Fig. 3 Throughput time. (a) (b) Comparison of AC Methods without failure Comparison of AC Methods with failure a Without machine failure, 45000 45000 b with machine failure 40000 40000 35000 35000 30000 30000 25000 25000 20000 20000 2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16 # machines n X n # machines n X n SRB2 Holonic DLRP SRB2 Holonic DLRP QLE LSIRP Ant QLE LSIRP Ant Fig. 4 Standard deviation of (a) (b) Comparison of AC Methods without failure Comparison of AC Methods with failure throughput time. a Without machine failure, b with machine failure 20000 15000 15000 10000 10000 5000 5000 0 0 2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16 # machines n X n # machines n X n SRB2 Holonic DLRP SRB2 Holonic DLRP QLE LSIRP Ant QLE LSIRP Ant increasing size of the production system, the SRB2 and Ant shows a different, worse performing behavior, significantly method cannot keep utilization on the same level. for bigger production systems with machine failures (see Regarding work-in-process (WIP), only the Ant method Fig. 6). mean throughput time deviation [sec] mean throughput time [sec] total output [items] mean throughput time deviation [sec] mean throughput time [sec] total output [items] Logist. Res. (2010) 2:109–120 119 Fig. 5 Machine utilization. Comparison of AC Methods without failure Comparison of AC Methods with failure (a) (b) a Without machine failure, 1 1 b with machine failure 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16 # machines n X n # machines n X n SRB2 Holonic DLRP SRB2 Holonic DLRP QLE LSIRP Ant QLE LSIRP Ant Fig. 6 Work-in-process. (a) Comparison of AC Methods without failure (b) Comparison of AC Methods with failure a Without machine failure, 20 20 b with machine failure 15 15 10 10 5 5 0 0 2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16 # machines n X n # machines n X n SRB2 Holonic DLRP SRB2 Holonic DLRP QLE LSIRP Ant QLE LSIRP Ant Although the simulation results cannot finally prove the helps delimiting different methods from each other apart behavior of autonomous control methods according to the from their actual implementation. The survey showed 14 presented pattern, they definitely support the assumption methods that have been developed in the recent years and that there is a systematic structure behind these methods. classified them according to the presented pattern. As a Two methods, namely Ant and SRB2, performed signifi- result, four different clusters containing similar methods cantly different regarding nearly all logistics target indi- could be identified. cators. Both methods are part of Cluster 1 and, at the same The simulation studies supported the assumption that time, do not belong to Cluster 3 (see Fig. 1). All other there are similarities in logistics performance between tested methods, however, are distributed among the other certain groups of autonomous control methods, while the clusters but behave similarly in the computer simulations. size of the production network does not significantly Another hint given by the simulation results is the auton- influence the methods’ behavior. omous control methods’ behavior as a function of pro- The research group will extend the simulation studies in duction network size. For all methods and all logistics the future in order to gain deeper knowledge on the relation target indicators, a relatively robust outcome can be between autonomous control method characteristics and observed. That means that in the presented scenario, the the logistics performance. Furthermore, it will focus on size of the production network does not have a major methods not included in the identified clusters, containing impact on the performance of the autonomous control the so far not implemented characteristics. methods. Acknowledgments This research is funded by the German Research Foundation (DFG) as part of the Collaborative Research Center 637 Autonomous Cooperating Logistic Processes—A Para- 5 Conclusion digm Shift and its Limitations (SFB 637) at Bremen University and at Jacobs University Bremen. This paper deals with the decision-making algorithms in autonomous control. A definition of the term autonomous References control method was presented followed by a classification pattern for autonomous control methods in logistics. The 1. Armbruster D, de Beer C, Freitag M, Jagalski T, Ringhofer C pattern offers the possibility to identify the components (2006) Autonomous control of production networks using a that make up a certain autonomous control method and pheromone approach. Physica A 363(1):104–114 work in progress [items] mean utilization mean utilization work in progress [items] 120 Logist. Res. (2010) 2:109–120 ¨ ¨ 2. Bose F, Windt K (2007) Catalogue of criteria for autonomous 14. Scholz-Reiter B, Gorges M, Philipp T (2009) Autonomously control. In: Hu ¨ lsmann M, Windt K (eds) Understanding autono- controlled production systems—influence of autonomous control mous cooperation and control in logistics—the impact on man- level on logistic performance. CIRP Ann Manuf Technol agement, information and communication and material flow. 58(1):395–398 Springer, Berlin, pp 57–72 15. Scho ¨ nberger J (2010) Adaptive demand peak management in 3. Bozarth CC, Warsing DP, Flynn BB, Flynn EJ (2009) The impact online transport process planning. OR Spectrum online first, of supply chain complexity on manufacturing plant performance. pp 1–29 J Oper Manag 27(1):78–93 16. Tsutsui S, Pelikan M (2007) cAS: the cunning ant system. 4. Cicirello V, Smith S (2001) Ant colony control for autonomous MEDAL report no 2007007 decentralized shop floor routing. In: ISADS-2001, fifth interna- 17. Ueda K, Hatono I, Fujii N, Vaario J (2000) Reinforcement tional symposium on autonomous decentralized systems learning approaches to biological manufacturing systems. CIRP 5. Fynes B, Voss C, de Bu ´ rca S (2005) The impact of supply chain Ann-Manuf Technol 49(1):343–346 relationship dynamics on manufacturing performance. Int J Oper 18. Vachon S, Klassen RD (2002) An exploratory investigation of the Prod Manag 25(1):6–19 effects of supply chain complexity on delivery performance. 6. Markus A, Kis Vancza T, Monostori L (1996) A market approach IEEE Trans Eng Manag 49(3):218–230. doi:10.1109/TEM.2002. to holonic manufacturing. CIRP Ann-Manuf Technol 45(1):433– 803387 436 19. Van Brussel H, Wyns J, Valckenaers P, Bongaerts L, Peeters P 7. Nyhuis P, Wiendahl H (2008) Fundamentals of production (1998) Reference architecture for holonic manufacturing systems: logistics: theory, tools and applications. Springer, New York PROSA. Comput Ind 37(3):255–274 8. Perona M, Miragliotta G (2004) Complexity management and 20. Vollmer L (2000) Agentenbasiertes Auftragsmanagement mit supply chain performance assessment. A field study and a con- Hilfe von Preis-Liefertermin-Relationen. VDI-Verlag, Du ¨ sseldorf ceptual framework. Int J Prod Econ 90(1):103–115 21. Wenning BL, Rekersbrink H, Timm-Giel A, Go ¨ rg C, Scholz- 9. Rekersbrink H, Makuschewitz T, Scholz-Reiter B (2009) A dis- Reiter B (2007) Autonomous control by means of distributed tributed routing concept for vehicle routing problems. Logist Res routing. In: Hu ¨ lsmann M, Windt K (eds) Understanding autono- 1(1):45–52 mous cooperation and control in logistics the impact on man- 10. Scholz-Reiter B, Windt K, Freitag M (2004) Autonomous logistic agement information and communication and material flow. processes: new demands and first approaches. In: Monostori L Springer, Berlin, pp 325–335 (ed) Proceedings of the 37th CIRP international seminar on 22. Windt K, Becker T (2009) Applying autonomous control methods manufacturing systems. Hungarian Academy of Science. Buda- in different logistic processes—a comparison by using an pest, pp 357–362 autonomous control application matrix. In: Proceedings of the 11. Scholz-Reiter B, Freitag M, de Beer C, Jagalski T (2006) The 17th mediterranean conference on control and automation, influence of production network’s complexity on the performance Thessaloniki of autonomous control methods. In: Teti R (ed) Intelligent 23. Windt K, Jeken O (2009) Allocation flexibility—a new flexibility computation in manufacturing engineering 5. Proceedings of the type as an enabler for autonomous control in production logistics. 5th CIRP international seminar on computation in manufacturing In: 42nd CIRP conference on manufacturing systems, Grenoble engineering (CIRP ICME ’06), University of Naples, Naples, 24. Windt K, Bo ¨ se F, Philipp T (2008) Autonomy in production pp 317–320 logistics—identification, characterisation and application. Int J 12. Scholz-Reiter B, de Beer C, Freitag M, Jagalski T (2008a) Bio- Robotics CIM 24(4):572–578 inspired and pheromone-based shop-floor control. Int J Comput 25. Windt K, Becker T, Kolev I (2010) A generic implementation Integr Manuf 21(2):201–205 approach of autonomous control methods in production logistics. 13. Scholz-Reiter B, Jagalski T, Bendul J (2008b) Autonomous In: Proceedings of the 43rd CIRP international conference on control of a shop floor based on bee’s foraging behaviour. In: manufacturing systems, Vienna Haasis HD, Kreowski HJ, Scholz-Reiter B (eds) Dynamics in logistics. First international conference, LDIC 2007, Springer, Berlin, pp 415–423 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Logistics Research Springer Journals

A classification pattern for autonomous control methods in logistics

Loading next page...
 
/lp/springer-journals/a-classification-pattern-for-autonomous-control-methods-in-logistics-yCcbCT8jtA
Publisher
Springer Journals
Copyright
Copyright © 2010 by Springer-Verlag
Subject
Engineering; Engineering Economics, Organization, Logistics, Marketing; Logistics; Industrial and Production Engineering; Simulation and Modeling; Operation Research/Decision Theory
ISSN
1865-035X
eISSN
1865-0368
DOI
10.1007/s12159-010-0030-9
Publisher site
See Article on Publisher Site

Abstract

Logist. Res. (2010) 2:109–120 DOI 10.1007/s12159-010-0030-9 OR IGINAL PAPER A classification pattern for autonomous control methods in logistics • • • Katja Windt Till Becker Oliver Jeken Achim Gelessus Received: 9 February 2010 / Accepted: 15 May 2010 / Published online: 13 June 2010 Springer-Verlag 2010 Abstract Autonomous control in logistics enables single 1 Introduction logistics objects to control the production and transporta- tion process. This shift from central planning to decen- In the past few years, a change can be observed in the area tralized control in real-time offers many possibilities to of logistics. Technological developments and changing cope with highly dynamic and complex systems. The market conditions have resulted in rising complexity in algorithms that define the decision behavior of each production and consequently in logistics. These changes logistics object, autonomous control methods, play a key include an increasing number of product variants, faster role in the successful implementation of autonomous con- delivery times, and shorter product life cycles [7]. The trol in logistics systems. A transparent classification is complexity of nowadays logistics processes has significant needed in order to identify the basic elements these impact on the performance of logistics processes in terms methods consist of. This classification supports the evalu- of delivery time and delivery reliability [3, 5, 8, 18]. A new ation of autonomous control methods in terms of gain- approach to deal with complexity is to increase the level ing knowledge about which method characteristics are of autonomous control in logistics processes [10]. The responsible for a method’s performance. This paper defines Collaborative Research Center (CRC 637) ‘‘Autonomous what autonomous control methods are, works out their Cooperating Logistic Processes—A Paradigm Shift and its fundamental characteristics, presents multiple methods Limitations’’ in Bremen, Germany, aims at developing new developed so far, and compares these methods regarding methods in production planning and control (e.g. [12, 13, characteristics and performance. 17]) as well as in transportation control (e.g., [9, 15]) in order to overcome the obstacles created by today’s com- Keywords Autonomous control  Categorization  plexity and dynamics. Simulation studies have already Decentral  Logistics  Production planning and control shown that increasing the level of autonomous control improves the logistic performance [14]. At the same time, the availability of new information technologies such as any type of wireless communication (e.g. RFID), distrib- K. Windt  T. Becker (&)  O. Jeken uted computing, and computer miniaturization serves as an Global Production Logistics, School of Engineering and Science, enabler for autonomous control. The CRC focuses its work Jacobs University Bremen gGmbH, Campus Ring 1, on decentralized methods that have the ability to utilize 28759 Bremen, Germany given flexibility potentials in logistics processes [23], but e-mail: t.becker@jacobs-university.de also on gaining more knowledge on autonomously con- K. Windt trolled system behavior, e.g. identifying the limitations of e-mail: k.windt@jacobs-university.de autonomous control. A. Gelessus Several different autonomous control methods have Computational Laboratory for Analysis, Modeling, been developed and tested in simulation studies. These and Visualization, School of Engineering and Science, previously conducted studies have illustrated that auton- Jacobs University Bremen gGmbH, Campus Ring 1, omous control can realize a higher logistics target 28759 Bremen, Germany 123 110 Logist. Res. (2010) 2:109–120 achievement in comparison to conventional production as well as a scheme for characterization. From the above planning and control [9, 14, 25]. However, it remains mentioned description of autonomous control and its ele- unclear which basic characteristics make up autonomous ments follows this definition: control and what is their influence on the logistics system An autonomous control method is a generic algorithm performance. The goal of this work is to identify basic that describes how logistics objects render and exe- elements of autonomous control methods and their influ- cute decisions by their own. ence on performance so that future methods can be developed systematically by focusing on parameters that There are many different ways an autonomous control contribute to a higher achievement of logistics targets. method can operate. A simple method could allow each This paper provides an overview on the current under- semi-finished part to choose the next production step in a standing of autonomous control in logistics as well as, job-shop scenario by preferring that machine with the more specific, on the term autonomous control methods in lowest number of waiting items in front of a machine Sect. 2. By collecting the basic elements of these methods, [11]. Another example is a method that is inspired by a classification pattern is elaborated. In Sect. 3, a literature ants’ foraging behavior. It uses virtual pheromones, survey of currently available autonomous control methods emitted by the parts, which then indicate a preferred path is presented. The methods are classified in terms of the through the production [4]. The diversity of possible previously developed pattern, and a systematic similarity autonomous control methods leads to the questions of the analysis is conducted. Section 4 presents a comparison of methods’ individual performance and calls for a compar- the methods’ performance. The performance evaluation has ison of these methods. A previously developed Autono- been done using computer simulations. The conclusion mous Control Application Matrix is available to support brings together all insights and points out the next steps for the evaluation and comparison efforts [22]. A comparison research in this field. of simulation results is presented in Scholz-Reiter et al. [14], but only taking into consideration three different methods and aiming at showing the relation between 2 Autonomous control methods process complexity and level of autonomous control. Beside the evaluation of single methods, it is necessary to For getting a better understanding how autonomous control understand the structure of a method and to be able to works if applied in manufacturing and transportation identify similarities and differences in order to interpret environments or in computer simulations for scientific the comparison results. Additionally, the categorization purposes, it is necessary to consider the single elements presented in this paper can be used in the future to create a toolbox for autonomous control method development. that compose an autonomous controlled system. Autono- mous control itself is defined as follows: Upon availability of performance results for multiple methods, the performance can be connected to the method ‘‘Autonomous control in logistics systems is charac- characteristics and thus offers information which charac- terized by the ability of logistics objects to process teristics to consider when creating a method for a specific information, to render and to execute decisions on application purpose. their own.’’ [24] A recent approach for classifying autonomous control is The definition consists of different elements from dif- called the Catalogue of Criteria for Autonomous Control in ferent layers. On the one hand, there are the logistics sys- Logistics [2]. This catalog collects criteria from the deci- tem, the logistics objects, and the information, which are sion-making, information processing, and decision execu- structural elements that describe the environment in which tion. For each criterion, four possible properties are offered autonomous control takes place. Information processing for classifying the logistics system. The result of the and decision rendering and execution on the other hand are classification leads to a degree of autonomous control. This activities that characterize the way the flow of goods is degree describes—taking into consideration the thirteen controlled in such a system. Thus, two different layers can different criteria—how much a system is autonomously be observed in a logistics system: the logistics processes as controlled and thus makes different logistics systems well as the control methods that operate on the logistics comparable in terms of the usage of autonomous control. processes. In this paper, the focus is set on control methods For the purpose presented in this paper, namely making the that enable logistics objects to take their own decision in building blocks of autonomous control methods visible, contrast to a scheduled and centralized production planning this approach is not feasible as it only offers the degree of and control strategy. autonomous control as a single figure but is not able to A prerequisite for the survey presented in Sect. 3 is compare multiple methods regarding their basic elements having a definition of the term autonomous control method of their construction. 123 Logist. Res. (2010) 2:109–120 111 Coming from the methods that have been analyzed, autonomous control methods are kept simple and have a characterization of autonomous control methods can be very short information horizon, which corresponds to done by using seven dimensions that have been defined: having 1 planning step, but some algorithms additionally temporal data, planning steps, artificial values, communi- take into consideration what can happen two, three, or cation type, data scope, actor, and data storage. Each more steps after the pending decision. Artificial values dimension is separated in reasonable sections, e.g. planning are figures that are not extracted directly from the envi- steps can be described in discrete steps as 1, 2, 3, or more ronment but are generated by the algorithm itself. These planning steps. For categorization purposes, the dimension artificial values can be pheromones as presented in the values have been arranged in up to four groups. An over- above mentioned ant approach, but also virtual money view of dimensions, values, and descriptions can be found used in auction-theoretic approaches. The Communication in Table 1. type describes how communication is conducted by the The dimension Temporal data indicates whether the algorithm. It distinguishes between communication method uses data from the past, future, or both. Past data among the moving logistics objects (parts), among the are defined as any kind of figure that can be extracted fixed objects (machines), among all objects, or commu- from the environment, e.g. buffer inventory level of a nication with a central control entity. The communication machine, moving average of processing time at a itself can range from a simple data request (e.g. a part machine, etc. Future data can be planned data or esti- demands the current buffer inventory level from a mated values for the above mentioned figures. Caution machine) to extensive negotiations between agents (e.g. should be exercised at this point as estimated values are parts bidding in an auction for resources). The Data often determined by calculating averages from past data, scope provides information about the number of variables and thus, these kind of values have to be classified as used for decision making. The dimension Actor points past data. The number of Planning steps points out how out whether the parts, the machines, or a central entity deep the algorithm searches in a decision tree. Usually act as decision maker in the algorithm. An issue regarding this dimension is the fact that from the decision point of view, it is not important who takes the decision. Only the input parameters and the target system define Table 1 Autonomous control categorization dimensions the outcome of the decision. This dimension is never- Dimension Values Description theless part of the categorization, as it plays an important role when implementing an autonomous control method Temporal data Past Indicates whether the method uses data from the past, future in a real logistics environment. The hardware layout of Future (planned), or both an autonomous control solution is strongly influenced by Hybrid this characteristic. Finally, Data storage describes where Planning steps 1 Number of future steps (e.g. the figures are stored that form the basis for decision machines) the method considers making. Again, this dimension is closely related to the hardware implementation. More Artificial values No Usage of artificial values for decision making, e.g. virtual Static pheromones 3 Autonomous control methods survey Dynamic Communication Part-machine Communication and data exchange In this section, autonomous control methods presented in type between logistics objects or a Part–part central entity and logistics the literature are described, characterized in accordance Machine– objects with the above defined dimensions, and compared by per- machine forming a systematic similarity analysis. Central Data scope Low (1–2) Number of variables used for 3.1 Methods description decision making Medium (3–5) High ([5) Various methods for autonomous control in logistics are Actor Part Logistics object that actively available today. Some of them have been developed in decides Machine the course of the research of the CRC 637, while others Central were invented without explicitly naming them autono- Data storage Part Location of data storage mous control methods. The descriptions of each single Machine autonomous control method identified can be found in Central Table 2. 123 112 Logist. Res. (2010) 2:109–120 Table 2 Autonomous control methods survey including method name (short name in parentheses), key idea, basic algorithm description, and remarks Method Ant algorithm (Ant) Cunning ant system (C-Ant) Source Ant colony control for autonomous decentralized shop floor The cunning ant system by Tsutsui and Pelikan [16 ] routing by Cicirello and Smith [4] Key ideas Ants carry products Two different types of ants exist Different jobs have different types of pheromones c-Ants, cunning Ants base most of their path on previous paths Machines have pheromone concentrations d-Ants are conventional ants, who donate their path to the c-Ants Ants choose machines based on pheromone concentration Both locate pheromones at spots on their path Pheromones expire over time different pheromones will be overwritten Optional: ants sometimes choose machines randomly pheromones evaporate after some time Algorithm 1. Initially choose a random machine 1. Let in some d-ants and some c-ants 2. Avoid machines with different pheromones 2. Let the d-ants choose the paths with the higher pheromone concentration with a higher likelihood 3. Go to machines with same pheromones 3. Let already deposited pheromones expire at a constant rate 4. Let pheromones expire after some time 4. Let c-ants retrieve the paths of the d-ants and use them partially for their own paths 5. Increase pheromone concentration if ant has visited the machine 6. Overwrite old pheromones if the pheromone type of the ant is different 7. Update pheromone concentrations once ant is through 8. Optional: choose a random machine Remarks Long initiation time; Algorithm avoids stagnations and is thus more flexible than the Not flexible if job influx changes ordinary ant algorithm Method Pheromone approach (Ph) Bee foraging (Bee) Source Autonomous control of production networks using Autonomous control of a shopfloor based on bee’s foraging a pheromone approach by Armbruster et al. [1] behaviour by Scholz-Reiter et al. [13] Key ideas Average throughput time is used as a pheromone Bees indicate good food sources via dancing To model evaporation, only the last n throughout Food sources are measured by quality and quantity times are considered Number of relevant parts for the throughput time More bees go to the better sources is the rate of evaporation Algorithm 1. Calculate average throughput times for the last n parts 1. Choose the best food place 2. Go for the machines with the lowest average 2. Different food sources (machines) exist throughput times a. Advertise (dance) and send information aa. Number of recruited bees depends on the number of dances ab. The quality of the source depends on the length of the dance b. Just use but don’t recruit c. Abandon machine and join pool of unemployed bees Remarks Slow adjustment to change Long initiation time Not flexible if job influx changes Method Simple rule based 1 (SRB 1) Simple rule based 2 (SRB 2) Source Autonomous control of a shop floor based The Influence of Production Networks’ Complexity on the on bee’s foraging behaviour by Performance of Autonomous Control Methods Scholz-Reiter et al. [13] by Scholz-Reiter et al. [11] Key ideas Compares estimated waiting time at buffers Compares estimated waiting time at buffers Uses future events Uses data from past events Algorithm 1. Parts are rated in estimated processing time 1. When a part leaves a machine it sends information about the processing times 123 Logist. Res. (2010) 2:109–120 113 Table 2 continued Method Simple rule based 1 (SRB 1) Simple rule based 2 (SRB 2) 2. Current buffer levels are calculated as a sum 2. This information is used by the following parts of the estimated processing time to decide where to go next 3. Choose the machine with the 3. Parts choose the machine which provides the lowest lowest processing time buffer mean duration of waiting and processing for parts of the same type Remarks Useful with high number of machines Useful with high number of products; Changes more slowly compared to the previous method Method Queue length estimator (QLE) Due date (DD) Source Modeling and analysis of production logistics processes Modeling and analysis of production logistics processes based on biologically inspired strategies by J. Bendul based on biologically inspired strategies by J. Bendul (Master’s thesis, University of Bremen, Germany) (Master’s thesis, University of Bremen, Germany) Key ideas Computes and estimates buffer states Uses the queue length estimator (QLE) Part decides autonomously based on various factors Orders parts by earliest due date Algorithm 1. All buffer states of machines that can perform 1. After a part leaves a machine it chooses its next the next step are computed destination based on the QLE method 2. The part decides whether to switch to a different 2. Within the queue of parts to be processed the part production line based on processing times or setup with the most urgent due date is chosen times, using local information to be processed next 3. Parts compare their own estimated time with the estimated time of the parts in the buffers and takes the machine with the minimal time Remarks Similar to the simple rule methods above Similar to the simple rule methods above Method One logistics target per rule (OLTPR) Market based control (market) Source Developed in this research group Developed in this research group according to Vollmer [20] Key ideas Implement various rules at the machines and parts, where Virtual currency is introduced each rule tried to achieve a specific logistics target Can be easily extended with new rules to further Parts carry a shopping list of work that needs improve outcome to be done on them Each job needed for a part has a budget associated Distance traveled to the machine has a price Parts auction for access to the machine Shopping List can be altered during the production process Algorithm 1. Utilization: each machine send a stronger attraction 1. Parts with shopping list and budgets enter the signal as its buffer becomes less full production process 2. On time delivery: parts are prioritized by their due date 2. Parts bid on the machines on their shopping list 3. On time delivery: parts prefer machines with 3. Highest bidder gets access to the machine short throughput time 4. The parts bid according to the minimal price of the machine and the distance cost 5. Machines grant access for the parts, if they are the highest bidder Remarks The various rules have to be weighted appropriately It is not clear how to choose the budgets to achieve good performance Price levels on the machines are important for production activity Development of price levels might be usable to investigate overall state of the production process, thus provide macro data Production is more dynamic as shopping lists can be altered during the process 123 114 Logist. Res. (2010) 2:109–120 Table 2 continued Method Holonic manufacturing (Holonic) Bionic manufacturing system (Bio) Source According to van Brussel [19]; A market approach to holonic Reinforcement learning approaches to Biological Manufacturing manufacturing by Markus et al. [6] Systems by Ueda [17] Key ideas Two agents bargain over the next item to be processed Attraction fields dependent on the type of job exist Agents are machines and management Fields attract specific jobs Management and machines bargain for the jobs to do Jobs have DNA like information about what work needs to be done on them Management punishes machines for delays Machine have operating knowledge that evolves Machines bid to get jobs from management Algorithm 1. After a part leaves a machine it chooses its next 1. Machines(Robots) are attracted by fields destination based on the QLE method 2. Within the queue of parts to be processed the part 2. Parts send out fields, depending on the production with the most urgent due date is chosen process they require to be processed next Remarks Punishment have to be chosen carefully No specific algorithm provided Existence of social dilemma; machine decision Based on manufacturing processes that require robots might cause overall loss, but gain for machine and on the spot machines Requires a central authority Method Link-state internet routing protocol (LSIRP) DLRP (DLRP) Source Developed in this research group according Autonomous control by Means of Distributed Routing to Wenning et al. [21] by Wenning et al. [21] Key ideas Based on a link-state routing protocol Parts request a route from machines Each machine has a map of the entire facility Machines communicate best routes to a destination Parts can be sorted according to any rule Algorithm 1. A map of the facility and the connections between 1. Each machine is a knowledge broker machines is built/provided 2. Shortest paths are computed based on various chosen 2. Before a part decides on a best route it ask the criteria, generally using Dijkstra’s algorithm current machine about possible ways to reach the destination 3. As the situation changes (breakdowns, buffer states, 3. Each machine includes relevant information from new machines) only the changes are propagated its knowledge base and forwards it to its successors among the machines 4. To make sure that insignificant changes are not 4. The successors do the same and forward this information propagated there must be a lower threshold along the production chain 5. The request is propagated through the network until the destination (drain) is reached 6. Then the last broker (the drain or the last machine) sends a reply directly to the part with all the collected information 7. After receiving one or multiple probable paths the part decides on the better way to take Remarks The machines can self-organize if information about A timeout may be included to reduce the amount the part production cycle is included of waiting for possible paths The threshold may need to be scaled by the Can be computationally expensive local buffer level 3.2 Systematic comparison of autonomous control the similarity between two methods. For two autonomous k l kl methods control methods M and M , the similarity is labeled S . Similarities can be defined for each categorization dimen- A systematic comparison of two autonomous control sion, and the similarity for the i-th categorization dimen- kl kl methods can be achieved by a mathematical definition of sion is labeled S . The total similarity S between two 123 Logist. Res. (2010) 2:109–120 115 k l autonomous control methods M and M is defined as the It should be mentioned that the temporal data M and arithmetic mean of the categorization-based similarities data scope M can also be represented as nonexclusive data kl kl 1 kl pairs (past/future) and (low/high), respectively, which leads S : S ¼ S . The categorization-based similari- i i n i¼1 kl kl kl kl to a similar definition of S and S as for S . Thus, for 5 ties S are defined for each categorization dimension 1 5 4 out of 7 categorization dimensions, an identical similarity individually. kl definition is used. The similarity S for temporal data is given by the The classification of the autonomous control methods following similarity matrix: with respect to the newly defined categorization dimen- Past Hybrid Future sions (see Table 1) is shown in Table 3. So far, the Past 1:00:50:0 ð1Þ Bionic Manufacturing System method (Bio) is conceptual Hybrid 1:00:5 research only, and a classification is not possible at the Future 1:0 moment. For the other 13 autonomous control methods, a According to this definition, two methods are completely classification according to the categorization dimensions similar with respect to the temporal data if they use data is possible. These data are used to calculate the simi- kl k l from the same temporal range. Distinctions about the data larities S for all pairs (M , M ). The full similarity number, origin, and quality are not made in the present matrix is shown in Table 4. For symmetry reasons, only analysis. the upper triangle is shown, and the diagonal elements, kl The similarity S for planning steps is defined by a 4 which are always 100%, are omitted. The SRB1 method 9 4 similarity matrix: and the QLE method have a similarity of 100%. This does not mean that both methods are identical. The ð1Þð2Þð3Þð [ 3Þ classification of the autonomous control methods is based ð1Þ 1:00:50:333 0:0 on a rather rough scheme, and small differences in any ð2Þ 1:00:50:0 ð2Þ of the categorization dimensions are not observable in ð3Þ 1:00:333 the analysis. ð [ 3Þ 1:0 Data from Table 4 are used to identify clusters of similar The present analysis does not make a distinction for methods. A method cluster is defined as a set of autono- methods using more than 3 planning steps. kl mous control methods with all similarities S larger than kl The similarity S for artificial values is defined by the the threshold 71.43%, which corresponds to an identical following similarity matrix: classification for 5 out of 7 categorization dimensions. Four No Static Dynamic autonomous control method clusters could be identified No 1:00:00:0 (see Fig. 1 for a graphical representation). The categori- ð3Þ Static 1:00:5 zation values for the four clusters are shown in Table 5. Dynamic 1:0 The clusters 1, 2, and 3 overlap, and 5 methods belong to more than one cluster. This indicates that all 8 methods Autonomous control methods using artificial values (Ant, Bee, DD, OLTPR, Ph, QLE, SRB1, and SRB2) of the have no similarity to methods using no artificial values. clusters 1, 2, and 3 are similar, which is confirmed by Autonomous control methods using static artificial values inspection of the categorization values (Table 5). Common are defined as 50% similar to methods using dynamic features for the autonomous control methods of the clusters artificial values. 1, 2, and 3 are For the communication type, the four values are non- exclusive, and more than one communication type can be – Planning Step is 1 realized in an autonomous control method. The similarity – Part–Machine communication is used kl S for the communication type is defined as the arithmetic 4 – Data Scope is either low or medium mean of the similarities for all communication types: – Decisions (Actor) are made by parts 4 – Data Storage is on the machine k l 1 1 : M ¼ M kl kl kl 4j 4j S ¼ S with S ¼ ð4Þ 4 4j 4j k l The Pheromone Approach fulfills the similarity criterion to 0 : M 6¼ M 4j 4j j¼1 all methods of the clusters 1, 2, and 3, and the methods Ant, kl The similarity S for data scope is defined by the same Bee, DD, OLTPR, QLE, SRB1, and SRB2 can be regarded kl similarity matrix as S with the labels replaced by low, as modifications of the Pheromone Approach although medium, and high. most methods have a completely different derivation. The kl kl The similarities S and S for actor and data storage, methods LSIRP and DLRP create cluster 4, which deviates 6 7 kl respectively, follow the as definition as for S , but the from cluster 1, 2, and 3 in the following categorization arithmetic mean is created from three contributions. values 123 116 Logist. Res. (2010) 2:109–120 – Planning Step is larger than 3 – Machine–Machine communication is used – Data Scope is high – Decisions (Actor) are made by machines The methods C-Ant, Market, and Holonic cannot be assigned to any of the method clusters, and each method is a singular development so far. The cluster analysis shows that many of the autono- mous control methods developed so far bear much resemblance to each other despite their derivations from different origins. So far, rarely used categorization values could be the starting point for the development of new autonomous control methods that differ significantly from the existing ones, and the following features should be considered: – Planning Step is larger than 1 – Artificial Values are used – Communication is not restricted to Part–Machine – Data Scope is high – Decisions (Actor) is not restricted to parts – Storage is not restricted to machine 4 Simulation studies After classifying and analyzing the different methods, a comparison regarding their performance will be presented in the following. To compare different autonomous control methods, a generic and common production scenario is used. The shop floor consists of m parallel production lines where each production line comprises n machines M ij (i = 1, …, m, and j = 1, …, n). Each machine has an input buffer B with a maximum level of 40 items. K is the ij number of different products that can be produced by the production system. At the source, the raw materials for each product type enter the system. The input frequencies for the raw materials are expressed by three temporally shifted sinus- shaped functions. This input model assumes a continu- ously varying number of incoming orders and reflects the dynamics within the production system. It is assumed that the different products have different operation times on the machines. The priority rule applied at the machines is first-come-first-served (FCFS) if no other rule is applied by the respective autonomous control. As the scenario is related to the classical job-shop factory layout, all production lines are connected at every stage. Every line is able to process every kind of product, and the parts can switch to a different line at each production stage. The product’s choice of the next processing machine is determined by the applied autonomous con- trol method. Table 3 Autonomous control methods classification Ant C-Ant Ph Bee SRB 1 SRB 2 QLE Temporal data Past Past Past Past Future Past Future Planning steps 1 More 1 1 1 1 1 Artificial values Dynamic Dynamic No Static No No No Communication type Part-machine Part-part Part-machine Part-part, part-machine Part-machine Part-machine Part-machine Data scope Low High Medium Medium Low Low Low Actor Part Part Part Part Part Part Part Data storage Machine Part, machine Machine Machine Machine Machine Machine DD OLTPR Market Holonic Bio LSIRP DLRP Temporal data Future Hybrid Past Future — Past Past Planning steps 1 1 1 3 — More More Artificial values No No Dynamic Dynamic — No No Communication type Part-machine, part-part Part-machine, part-part Part-machine, central Central, machine-machine — Machine-machine, central Machine-machine, part-machine Data scope Medium Medium High High — High High Actor part Part Part, machine Part, machine, central Machine, central — Machine Part, machine Data storage Part, machine Part, machine Part, machine, central Machine, central — Machine, central Machine Logist. Res. (2010) 2:109–120 117 Table 4 Autonomous control methods similarity in % C-Ant Ph Bee SRB 1 SRB 2 QLE DD OLTPR Market Holonic LSIRP DLRP 60 79 82 71 86 71 56 58 63 36 32 49 Ant 52 63 31 45 31 46 49 61 45 56 65 C-Ant 82 79 93 79 77 80 56 29 54 70 Ph 61 75 61 67 69 60 32 36 52 Bee 86 100 85 73 35 36 32 49 SRB 1 86 70 73 49 21 46 63 SRB 2 85 73 35 36 32 49 QLE 88 43 35 31 48 DD 55 32 43 60 OLTPR 63 50 50 Market 54 37 Holonic 83 LSIRP Cluster 2 output quantity during the simulation run is counted. Fig- Cluster 4 ure 2 shows the results. Except for the Ant and SRB2 algorithm, the investigated methods show a linear increase LSIRP Ant SRB 2 SRB 1 in the output quantity for both cases with and without QLE DLRP Bee machine failures. This shows the general performance but Ph OLTPR does not take into consideration the achievement of the different logistics targets. The logistics targets are short DD lead times, high due date reliability, high utilization, and Cluster 1 Cluster 3 low inventory [7]. In the simulation, the lead time is measured as the mean throughput time of the parts in the manufacturing system. The due date reliability is measured Holonic C-Ant as the standard deviation of the throughput time, because it Market is assumed that low standard deviation indicates high predictability of lead time and results in high due date Fig. 1 Autonomous control method clusters performance. The utilization is measured as percentage of time the machines are busy in relation to the total time To compare the different methods, simulations were capacity. The inventory level is measured as the average conducted for production scenarios from m = 3to m = 9 number of parts in the buffers in front of the machines. with m = n and K = 3 (products A, B, and C). The Figure 3 shows the mean throughput time for the dif- operation time was determined in relation to the simulation ferent methods. Without machine failures, the distribution time. Product A had an operation time of 0.1%, product B appears evenly among the methods, having QLE, Holonic, of 0.5%, and product C of 1% of the total simulation time. LSIRP, and DLRP performing in a similar pattern, whereas To analyze the performance of different control methods, Ant and SRB2 perform significantly worse (see Fig. 3a). In all methods that have been described in detail by their case of machine failures, SRB2 still performs worse than publication sources were selected. As the autonomous the other methods, but the difference compared to the other control methods are designed to cope with complex and methods is smaller as displayed in Fig. 3b. The standard dynamic manufacturing systems, the comparison is carried deviation of the throughput time shows again the distinct out for scenarios with and without machine breakdowns. difference between the SRB2 and Ant method in relation to The breakdowns of the machines represent an additional the rest of the investigated methods (see Fig. 4). However, kind of dynamics and unpredictability of the system beside again, the difference is smaller in the case of machine the input fluctuation. The breakdowns occurred randomly failures than that without machine failures. The two with a probability of 75% and a repair time of 1 min. machine utilization graphs (see Fig. 5) show a similar In order to determine the performance of the different pattern of performance where the drop in the performance autonomous control methods, several key indicators are in Fig. 5b is caused by the machine failure. The previously measured. All results presented here are average values mentioned well-performing group of methods achieves to over five independent simulation runs. First, the overall keep the utilization on a high level, whereas with an 123 118 Logist. Res. (2010) 2:109–120 Table 5 Common characteristics of the method clusters identified Methods Cluster 1 Cluster 2 Cluster 3 Cluster 4 Ant, Ph, Bee, SRB 2 Ph, SRB 1, SRB 2, QLE, OLTPR Ph, SRB 1, QLE, DD, OLTPR LSIRP, DLRP Temporal data Past Different values Different values Past Planning steps 1 1 1 [3 Artificial values Different values No No No Communication type Always including Always including Always including Always including part-machine part-machine part-machine machine-machine Data scope Low/medium Low/medium Low/medium High Actor Part Always including part Always including part Always including machine Data storage Machine Always including machine Always including machine Always including machine Fig. 2 Total output of Comparison of AC Methods without failure Comparison of AC Methods with failure (a) (b) simulation run. a Without 3000 3000 machine failure, b with machine 2500 2500 failure 1500 1500 1000 1000 500 500 0 0 2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16 # machines n X n # machines n X n SRB2 Holonic DLRP SRB2 Holonic DLRP QLE LSIRP Ant QLE LSIRP Ant Fig. 3 Throughput time. (a) (b) Comparison of AC Methods without failure Comparison of AC Methods with failure a Without machine failure, 45000 45000 b with machine failure 40000 40000 35000 35000 30000 30000 25000 25000 20000 20000 2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16 # machines n X n # machines n X n SRB2 Holonic DLRP SRB2 Holonic DLRP QLE LSIRP Ant QLE LSIRP Ant Fig. 4 Standard deviation of (a) (b) Comparison of AC Methods without failure Comparison of AC Methods with failure throughput time. a Without machine failure, b with machine failure 20000 15000 15000 10000 10000 5000 5000 0 0 2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16 # machines n X n # machines n X n SRB2 Holonic DLRP SRB2 Holonic DLRP QLE LSIRP Ant QLE LSIRP Ant increasing size of the production system, the SRB2 and Ant shows a different, worse performing behavior, significantly method cannot keep utilization on the same level. for bigger production systems with machine failures (see Regarding work-in-process (WIP), only the Ant method Fig. 6). mean throughput time deviation [sec] mean throughput time [sec] total output [items] mean throughput time deviation [sec] mean throughput time [sec] total output [items] Logist. Res. (2010) 2:109–120 119 Fig. 5 Machine utilization. Comparison of AC Methods without failure Comparison of AC Methods with failure (a) (b) a Without machine failure, 1 1 b with machine failure 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0 0 2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16 # machines n X n # machines n X n SRB2 Holonic DLRP SRB2 Holonic DLRP QLE LSIRP Ant QLE LSIRP Ant Fig. 6 Work-in-process. (a) Comparison of AC Methods without failure (b) Comparison of AC Methods with failure a Without machine failure, 20 20 b with machine failure 15 15 10 10 5 5 0 0 2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16 # machines n X n # machines n X n SRB2 Holonic DLRP SRB2 Holonic DLRP QLE LSIRP Ant QLE LSIRP Ant Although the simulation results cannot finally prove the helps delimiting different methods from each other apart behavior of autonomous control methods according to the from their actual implementation. The survey showed 14 presented pattern, they definitely support the assumption methods that have been developed in the recent years and that there is a systematic structure behind these methods. classified them according to the presented pattern. As a Two methods, namely Ant and SRB2, performed signifi- result, four different clusters containing similar methods cantly different regarding nearly all logistics target indi- could be identified. cators. Both methods are part of Cluster 1 and, at the same The simulation studies supported the assumption that time, do not belong to Cluster 3 (see Fig. 1). All other there are similarities in logistics performance between tested methods, however, are distributed among the other certain groups of autonomous control methods, while the clusters but behave similarly in the computer simulations. size of the production network does not significantly Another hint given by the simulation results is the auton- influence the methods’ behavior. omous control methods’ behavior as a function of pro- The research group will extend the simulation studies in duction network size. For all methods and all logistics the future in order to gain deeper knowledge on the relation target indicators, a relatively robust outcome can be between autonomous control method characteristics and observed. That means that in the presented scenario, the the logistics performance. Furthermore, it will focus on size of the production network does not have a major methods not included in the identified clusters, containing impact on the performance of the autonomous control the so far not implemented characteristics. methods. Acknowledgments This research is funded by the German Research Foundation (DFG) as part of the Collaborative Research Center 637 Autonomous Cooperating Logistic Processes—A Para- 5 Conclusion digm Shift and its Limitations (SFB 637) at Bremen University and at Jacobs University Bremen. This paper deals with the decision-making algorithms in autonomous control. A definition of the term autonomous References control method was presented followed by a classification pattern for autonomous control methods in logistics. The 1. Armbruster D, de Beer C, Freitag M, Jagalski T, Ringhofer C pattern offers the possibility to identify the components (2006) Autonomous control of production networks using a that make up a certain autonomous control method and pheromone approach. Physica A 363(1):104–114 work in progress [items] mean utilization mean utilization work in progress [items] 120 Logist. Res. (2010) 2:109–120 ¨ ¨ 2. Bose F, Windt K (2007) Catalogue of criteria for autonomous 14. Scholz-Reiter B, Gorges M, Philipp T (2009) Autonomously control. In: Hu ¨ lsmann M, Windt K (eds) Understanding autono- controlled production systems—influence of autonomous control mous cooperation and control in logistics—the impact on man- level on logistic performance. CIRP Ann Manuf Technol agement, information and communication and material flow. 58(1):395–398 Springer, Berlin, pp 57–72 15. Scho ¨ nberger J (2010) Adaptive demand peak management in 3. Bozarth CC, Warsing DP, Flynn BB, Flynn EJ (2009) The impact online transport process planning. OR Spectrum online first, of supply chain complexity on manufacturing plant performance. pp 1–29 J Oper Manag 27(1):78–93 16. Tsutsui S, Pelikan M (2007) cAS: the cunning ant system. 4. Cicirello V, Smith S (2001) Ant colony control for autonomous MEDAL report no 2007007 decentralized shop floor routing. In: ISADS-2001, fifth interna- 17. Ueda K, Hatono I, Fujii N, Vaario J (2000) Reinforcement tional symposium on autonomous decentralized systems learning approaches to biological manufacturing systems. CIRP 5. Fynes B, Voss C, de Bu ´ rca S (2005) The impact of supply chain Ann-Manuf Technol 49(1):343–346 relationship dynamics on manufacturing performance. Int J Oper 18. Vachon S, Klassen RD (2002) An exploratory investigation of the Prod Manag 25(1):6–19 effects of supply chain complexity on delivery performance. 6. Markus A, Kis Vancza T, Monostori L (1996) A market approach IEEE Trans Eng Manag 49(3):218–230. doi:10.1109/TEM.2002. to holonic manufacturing. CIRP Ann-Manuf Technol 45(1):433– 803387 436 19. Van Brussel H, Wyns J, Valckenaers P, Bongaerts L, Peeters P 7. Nyhuis P, Wiendahl H (2008) Fundamentals of production (1998) Reference architecture for holonic manufacturing systems: logistics: theory, tools and applications. Springer, New York PROSA. Comput Ind 37(3):255–274 8. Perona M, Miragliotta G (2004) Complexity management and 20. Vollmer L (2000) Agentenbasiertes Auftragsmanagement mit supply chain performance assessment. A field study and a con- Hilfe von Preis-Liefertermin-Relationen. VDI-Verlag, Du ¨ sseldorf ceptual framework. Int J Prod Econ 90(1):103–115 21. Wenning BL, Rekersbrink H, Timm-Giel A, Go ¨ rg C, Scholz- 9. Rekersbrink H, Makuschewitz T, Scholz-Reiter B (2009) A dis- Reiter B (2007) Autonomous control by means of distributed tributed routing concept for vehicle routing problems. Logist Res routing. In: Hu ¨ lsmann M, Windt K (eds) Understanding autono- 1(1):45–52 mous cooperation and control in logistics the impact on man- 10. Scholz-Reiter B, Windt K, Freitag M (2004) Autonomous logistic agement information and communication and material flow. processes: new demands and first approaches. In: Monostori L Springer, Berlin, pp 325–335 (ed) Proceedings of the 37th CIRP international seminar on 22. Windt K, Becker T (2009) Applying autonomous control methods manufacturing systems. Hungarian Academy of Science. Buda- in different logistic processes—a comparison by using an pest, pp 357–362 autonomous control application matrix. In: Proceedings of the 11. Scholz-Reiter B, Freitag M, de Beer C, Jagalski T (2006) The 17th mediterranean conference on control and automation, influence of production network’s complexity on the performance Thessaloniki of autonomous control methods. In: Teti R (ed) Intelligent 23. Windt K, Jeken O (2009) Allocation flexibility—a new flexibility computation in manufacturing engineering 5. Proceedings of the type as an enabler for autonomous control in production logistics. 5th CIRP international seminar on computation in manufacturing In: 42nd CIRP conference on manufacturing systems, Grenoble engineering (CIRP ICME ’06), University of Naples, Naples, 24. Windt K, Bo ¨ se F, Philipp T (2008) Autonomy in production pp 317–320 logistics—identification, characterisation and application. Int J 12. Scholz-Reiter B, de Beer C, Freitag M, Jagalski T (2008a) Bio- Robotics CIM 24(4):572–578 inspired and pheromone-based shop-floor control. Int J Comput 25. Windt K, Becker T, Kolev I (2010) A generic implementation Integr Manuf 21(2):201–205 approach of autonomous control methods in production logistics. 13. Scholz-Reiter B, Jagalski T, Bendul J (2008b) Autonomous In: Proceedings of the 43rd CIRP international conference on control of a shop floor based on bee’s foraging behaviour. In: manufacturing systems, Vienna Haasis HD, Kreowski HJ, Scholz-Reiter B (eds) Dynamics in logistics. First international conference, LDIC 2007, Springer, Berlin, pp 415–423

Journal

Logistics ResearchSpringer Journals

Published: Jun 13, 2010

References