This paper presents an analysis of the trends and behavior of Fitness Landscape Analysis (FLA) and corresponding algorithm performance features for instances of the Quadratic Assignment Problem (QAP) and the instance space between them. Given two QAPLIB instances, a transformation generates 30 intermediary instances, i.e. problem versions for further experimentation. For each problem version, we track algorithm performance of robust tabu search (RTS) and variable neighborhood search (VNS), as well as FLA measures obtained by various types of walks. Thus, we are able to analyze how these performances and measures change during the transformation. We observe that RTS dominates VNS in earlier problem versions, while VNS outperforms RTS in later problem versions. Overall, the transformation leads to a smooth traversal of the instance space, and both algorithm performance and FLA measures correlate with problem versions.

Industrial manufacturing is currently amidst it’s fourth great revolution, pushing towards the digital transformation of production processes. One key element of this transformation is the formalization and digitization of processes, creating an increased potential to monitor, understand and optimize existing processes. However, one major obstacle in this process is the increased diversification and specialisation, resulting in the dependency on multiple experts, which are rarely amalgamated in small to medium sized companies. To mitigate this issue, this paper presents a novel approach for multi-criteria optimization of workflow-based assembly tasks in manufacturing by combining a workflow modeling framework and the HeuristicLab optimization framework. For this endeavour, a new generic problem definition is implemented in HeuristicLab, enabling the optimization of arbitrary workflows represented with the modeling framework. The resulting Pareto front of the multi-criteria optimization provides the decision makers a set of optimal workflows from which they can choose to optimally fit the current demands. The advantages of the herein presented approach are highlighted with a real world use case from an ongoing research project.

The performance of modern production systems often depends upon automated production planning strategies such as material requirements planning. Parametrizing, evaluating and comparing these strategies by testing them in the real world is often difficult and prohibitively resource intensive. State-of-the-art computer simulation can be used to adequately model the production processes and predict the relevant performance metrics without investing valuable production capacities. Heuristic optimization procedures can build on these simulations to fine-tune production planning strategies. A major obstacle for this simulation-based optimization approach, however, lies in its computational requirements since accurate production simulations require their fair share of computation time. In this work, we will demonstrate the use of heuristic optimization to learn optimal production strategies for a bi-objective high-dimensional real world scenario and explore how to reduce the computational cost of the heuristic search by use of surrogates and dimensionality reduction. Results indicate that the employed approach achieved solutions that could outperform the production planning parameters currently used in the real world.

Changes in dynamic optimization problems entail updates to the problem model, which in turn can result in changes to the problem’s fitness landscape and even its solution encoding. In order to yield valid solutions that are applicable to the current problem state, optimization algorithms must be able to cope with such dynamic problem updates. Furthermore, depending on the optimization use case, changes occurring in real-world environments require an optimizer to adapt to changing process conditions and yield updated, valid solutions within a short time frame. In this paper, dynamic problem changes and their effects on an optimizer’s algorithmic behavior are studied in the context of crane scheduling. Three open-ended versions of RAPGA, a relevant alleles preserving genetic algorithm, are evaluated, some of which include self-adaption and a special treatment of certain events that require domain knowledge to be recognized. The proposed extensions affect the algorithm behavior as desired. On the one hand, the algorithms converge faster after a loss in solution quality is detected. On the other hand, new genetic material is introduced, making it possible to reach high quality areas of the search space again.

In dynamic warehouse operations, factory crane scheduling is a challenging problem to be solved. The interplay between a number of cranes requires good coordination to efficiently handle requested transport orders within the warehouse. The objective is to maximize the throughput and therefore minimize the makespan of crane schedules for a given set of transport orders and resulting crane moves. However, in real world processes, uncertainty is a key factor that needs to be considered. New transport orders arrive without prior announcement during the planning window, as the exact timing of vehicles responsible for inbound and outbound flows are often available only on short notice. Furthermore, events such as crane maintenance or operator breaks must be dealt with accordingly. To further the algorithmic research where these challenging events in the domain of dynamic factory crane scheduling must be tackled, we propose an open-source simulation model that serves as a novel benchmark.

Dynamic optimization problems (DOPs) are an underrepresented class in benchmarking evolutionary computation systems (ECS). Most benchmarks focus on more or less expensive problems, but which never change during the optimization. In real-world logistics operations however, dynamic changes and even uncertainty are natural and have to be dealt with. While evolutionary algorithms are certainly well suited methods to tackle such problems, the field lacks public and open source, easy-to-use, but still complex dynamic environments for comparing and further developing the methods. In this work, we highlight the framework that we have created and open sourced as part of the DynStack competition which was first held at GECCO 2020. We present the underlying principles of the framework, the architecture that eases the application, and potential ways to benchmark a range of methods. The environments implemented in this framework are real-world industrial scenarios, that have been simplified, but which still convey practical challenges in the application of ECS to real-world problems.

Many real-world processes are of dynamic nature and therefore subject to change. In this paper, dynamic warehouse operations are taken care of, more specifically crane operations that involve moving steel coils between storage locations within a large warehouse. An open-ended optimization approach is employed to create an optimal schedule of crane moves given a set of requested crane operations. Conventionally, the problem model defines static crane speeds and service times, the time needed to pickup and dropoff coils from/to locations. In a dynamic environment, these properties can depend on a variety of factors, including the proficiency of the crane operator or the storage locations that are accessed. Therefore, an open-ended genetic algorithm is enhanced with integrated machine learning (IML) tasked with learning crane speeds and service times from historical data and adapting said properties in the underlying problem model in order to provide the optimizer with a more realistic view on the current world state. To understand the performance gain achieved by this enhancement, experimental setups with and without IML are evaluated. The results show that IML improves the optimizer’s performance, as the algorithm gains better understanding of the current world state and is therefore able to create more suitable schedules, considering the crane’s current performance.

Dynamic optimization problems (DOPs) are an underrepresented class in benchmarking evolutionary computation systems (ECS). Most benchmarks focus on more or less expensive problems, but which never change during the optimization. In real-world logistics operations however, dynamic changes and even uncertainty are natural and have to be dealt with. While evolutionary algorithms are certainly well suited methods to tackle such problems, the field lacks public and open source, easy-to-use, but still complex dynamic environments for comparing and further developing the methods. In this work, we highlight the framework that we have created and open sourced as part of the DynStack competition which was first held at GECCO 2020. We present the underlying principles of the framework, the architecture that eases the application, and potential ways to benchmark a range of methods. The environments implemented in this framework are real-world industrial scenarios, that have been simplified, but which still convey practical challenges in the application of ECS to real-world problems.

Fitness Landscape Analysis (FLA) denotes the task of analyzing black-box optimization problems and capturing their characteristic features with the goal of providing additional information, that may help in algorithm selection, parametrization or guidance. Many real-world optimization tasks require dynamic on-going optimization and a plethora of meta-heuristic algorithms has been introduced for this task. However, most analysis focuses on static problems or dynamic optimization tasks without time-linkage, where the dynamic changes of the problem are independent of the decisions taken by the optimizer, but many real-world optimization problems display very heavy dependence on previous states and decisions. In this paper, the techniques of the static FLA are combined with dynamic and domain specific measures and applied to two dynamic problems. A time-linked dynamic OneMax problem and a dynamic multi-objective knapsack problem are presented and the impact of time-linkage on their FLA features is analyzed.

Dynamic optimization is of high practical relevance for many production and logistics processes. Often however, in research, the dynamics are neglected and an algorithm or optimization is presented for a static decision scenario. The effects that occur with implementing decisions one by one in a dynamic environment subject to other dynamic events have not received as much attention as static problem scenarios. We think this is in part due to the increased complexity of describing a dynamic environment and parameterizing it in a meaningful way. In this publication we present three dynamic environments in the context of warehouse logistics where manipulation occurs by gantry cranes. Our scenarios are heavily inspired by real-world steel logistics operations. We provide free open source implementations of these dynamic environments.

Solving manufacturing optimization problems in the context of intelligent production involves the consideration of continuously changing events of the respective enterprise environment in real time. Smart solution methods are needed which are able to cope with such necessary reactions to uncertainty and dynamics. In general, this field of research belongs to the topic of dynamic optimization. However, investigating the relevant literature reveals the broad range of this research area. In addition to real time, i.e. online optimization it contains a large number of other (dynamic) sectors. After differentiating dynamic online optimization from other research domains of dynamic optimization, the aim of this work is (1) to show in which streams and problem fields it has already been investigated, and (2) which different approaches to categorize online optimization problems are known so far. As a result, an overview of the state of the art concerning the occurence and existing categorizations of online optimization problems in the context of smart manufacturing is given, demonstrating ambiguities in the language used and in the categorization efforts for this optimization problem and therefore motivating further research efforts on a comprehensive integration of the findings of different streams in this area.

Fitness landscape analysis (FLA) is a useful tool in the domain of (meta-)heuristic optimization but depends on explicitly knowing what fitness value is assigned to each solution. Dynamic optimization problems often do not provide their fitness landscape in such an explicit form, but by employing problem-specific knowledge, information about the problem itself and its current state can still be obtained. In this paper, a type of gray-box analysis of states of the open-ended stacking problem in two variations is presented. The current states obtained by monitoring the problem and algorithm during optimization are described via statistical measures similar to FLA measures. From this the distribution of possible states (the state landscape) and the transitions between problem states are analyzed. Visualization of the empirically obtained results reveals insights into algorithm-problem dynamics.

Worker cross-training is a problem arising in many companies that involve human work. To perform certain activities, workers are required to possess certain skills. Cross-trained workers possess ev...

Efficient global optimization is, even after over two decades of research, still considered as one of the best approaches to surrogate-assisted optimization. In this paper, material requirements planning parameters are optimized and two different versions of EGO, implemented as optimization networks in HeuristicLab, are applied and compared. The first version resembles a more standardized version of EGO, where all steps of the algorithm, i.e. expensive evaluation, model building and optimizing expected improvement, are executed synchronously in sequential order. The second version differs in two aspects: (i) instead of a single objective, two objectives are optimized and (ii) all steps of the algorithm are executed asynchronously. The latter leads to faster algorithm execution, since model building and solution evaluations can be done in parallel and do not block each other. Comparisons are done in terms of achieved solution quality and consumed runtime. The results show that the multi-objective, asynchronous optimization network can compete with the single-objective, synchronous version and outperforms the latter in terms of runtime.

Worker cross-training is a problem arising in many industries and companies that involve human work, since workers that possess multiple skills, i.e., a qualification profile, may be employed more flexible on a day-to-day basis. At the same time it can be assumed that these workers are also incur a higher personnel cost. It is therefore of high interest to a company to balance the available skills such that customer deadlines can be met in a cost-efficient way. In this work we extend a simulation-based optimization approach with a third objective and apply NSGA-II.

The dynamic block relocation problem is a variant of the BRP where the initial configuration and retrieval priorities are known but are subject to change during the implementation of an optimized solution. This paper investigates two kinds of potential changes. The exchange of assigned priorities between two blocks and the arrival of new blocks. For both kind of events we present algorithms that can adjust an existing solution to the changed situation. These algorithms are combined with a branch and bound based solver to enable online optimization with look-ahead. Our experiments show that the algorithms enable finding better solutions in a shorter time after a event occurs.

To react on increasing customer demand uncertainty, production systems have to be flexible concerning the provided capacity. With respect to labour, one opportunity to gain such flexibility is to assign workers to different work stations which often require different skills to be operated. Therefore, cross-trained workers are needed to enable this flexibility. Since the qualifying workforce implies costs, a relevant problem is how much skills and what mix of skills is optimal for a production system. In addition, the workforce may be on vacation or have a sick leave and hence is not always available. In this paper, we study the effect of different predefined workforce qualification profiles for a streamlined production system with simulation and compare the results with simulation-based optimisation using a genetic algorithm. Specifically the effect of stochastic worker absence, in comparison to workers being always available, is evaluated for different production system scenarios. The results show that cross-trained workers can significantly improve the service level achieved and that simulation-based optimisation can provide a much better worker specific mix of skills than predefined qualification profiles. Another managerial insight is that there is a trade-off between number of skills and number of workers needed to obtain the same service level.

In this paper we present a dynamic stacking problem with uncertainty. We developed a simulation environment, an optimizer for solving it, and performance measures to determine the success of the optimizer. The problem requires handling incoming blocks, stacking them efficiently, and meeting deadlines for delivery, while not knowing exactly when blocks will arrive or when they will be ready for delivery. The optimizer models the problem as a dynamic Block Relocation Problem and solves it using a branch&bound based heuristic. The simulation and optimizer run concurrently and the distribution of random variables is not disclosed to the solver and must, therefore, be estimated. We study the influence of uncertainty on the solver and show that the degree of uncertainty has a significant impact on the performance of the overall system. We also experiment with different measures to estimate uncertain arrival times and show that the choice of measure is important for achieving good performance.

In the context of real-world optimization problems in the area of production and logistics, multiple objectives have to be considered very often. Precisely such a situation is also regarded in this work. For a resource-constrained project scheduling problem with activity selection and time flexibility, a new bi-objective extension is developed. Motivated by a steel industry production case, each of two already existing objective functions, makespan minimization and time balance maximization, is deployed together with a newly developed objective, the so-called activity priority maximization. To solve the resulting two bi-objective optimization problems and provide all existing trade-off solutions, the ϵ -constraint method is used. A constraint programming model is presented and solved with the CP Optimizer of IBM ILOG CPLEX and the results are compared concerning solution quality and runtime, showing the competitiveness of the developed model.

Project scheduling in manufacturing environments often requires flexibility in terms of the selection and the exact length of alternative production activities. Moreover, the simultaneous scheduling of multiple lots is mandatory in many production planning applications. To meet these requirements, a new resource-constrained project scheduling problem (RCPSP) is introduced where both decisions (activity flexibility and time flexibility) are integrated. Besides the minimization of makespan, two new alternative objectives are presented: maximization of balanced length of selected activities (time balance) and maximization of balanced resource utilization (resource balance). New mixed integer and constraint programming (CP) models are proposed for the developed integrated flexible project scheduling problem. Benchmark instances on an already existing flexible RCPSP and the newly developed problem are solved to optimality. The real-world applicability of the suggested CP models is shown by additionally solving a large industry case.