Johannes Karder

Towards the design and implementation of optimization networks in HeuristicLab

Combining multiple algorithms to cooperate in solving different optimization problems or process other workflows can be done in various problem domains, e.g. combinatorial optimization and data analysis. Optimization networks allow us to create such cooperative approaches by connecting multiple algorithms and letting them work together. In this paper, we propose an optimization network architecture for HeuristicLab. Networks are built using nodes that perform arbitrary tasks. We introduce the concepts of messages and ports, which can be used to exchange data between nodes. The application of such optimization networks is shown for two different applications. One is to solve the Traveling Thief Problem, where we substitute parts of the original problem with subproblems that are optimized interdependently. In another scenario, feature selection is combined with linear regression to find the best combination of features in order to achieve the best linear regression model.

Solving the Traveling Thief Problem Using Orchestration in Optimization Networks.

Optimization problems can sometimes be divided into multiple subproblems. Working on these subproblems instead of the actual master problem can have some advantages, e.g. if they are standard problems, it is possible to use already existing algorithms, whereas specialized algorithms would have to be implemented for the master problem. In this paper we approach the NP-hard Traveling Thief Problem by implementing different cooperative approaches using optimization networks. Orchestration is used to guide the algorithms that solve the respective subproblems. We conduct experiments on some instances of a larger benchmark set to compare the different network approaches to best known results, as well as a sophisticated, monolithic approach. Using optimization networks, we are able to find new best solutions for all of the selected problem instances.

Scripting and framework integration in heuristic optimization environments

Rapid prototyping and testing of new ideas has been a major argument for evolutionary computation frameworks. These frameworks facilitate the application of evolutionary computation and allow experimenting with new and modified algorithms and problems by building on existing, well tested code. However, one could argue, that despite the many frameworks of the metaheuristics community, software packages such as MATLAB, GNU Octave, Scilab, or RStudio are quite popular. These software packages however are associated more closely with numerical analysis rather than evolutionary computation. In contrast to typical evolutionary computation frameworks which provide standard implementations of algorithms and problems, these popular frameworks provide a direct programming environment for the user and several helpful functions and mathematical operations. The user does not need to use traditional development tools such as a compiler or linker, but can implement, execute, and visualize his ideas directly within the environment. HeuristicLab has become a popular environment for heuristic optimization over the years, but has not been recognized as a programming environment so far. In this article we will describe new scripting capabilities created in HeuristicLab and give information on technical details of the implementation. Additionally, we show how the programming interface can be used to integrate further metaheuristic optimization frameworks in HeuristicLab. Categories and Subject D.

Asynchronous surrogate-assisted optimization networks

This paper introduces a new, highly asynchronous method for surrogate-assisted optimization where it is possible to concurrently create surrogate models, evaluate fitness functions and do parameter optimization for the underlying problem, effectively eliminating sequential workflows of other surrogate-assisted algorithms. Using optimization networks, each part of the optimization process is exchangeable. First experiments are done for single objective benchmark functions, namely Ackley, Griewank, Schwefel and Rastrigin, using problem sizes starting from 2D up to 10D, and other EGO implementations are used as baseline for comparison. First results show that the implemented network approach is competitive to other EGO implementations in terms of achieved solution qualities and more efficient in terms of execution times.

Optimization Networks for Integrated Machine Learning

Optimization networks are a new methodology for holistically solving interrelated problems that have been developed with combinatorial optimization problems in mind. In this contribution we revisit the core principles of optimization networks and demonstrate their suitability for solving machine learning problems. We use feature selection in combination with linear model creation as a benchmark application and compare the results of optimization networks to ordinary least squares with optional elastic net regularization. Based on this example we justify the advantages of optimization networks by adapting the network to solve other machine learning problems. Finally, optimization analysis is presented, where optimal input values of a system have to be found to achieve desired output values. Optimization analysis can be divided into three subproblems: model creation to describe the system, model selection to choose the most appropriate one and parameter optimization to obtain the input values. Therefore, optimization networks are an obvious choice for handling optimization analysis tasks.

A General Solution Approach for the Location Routing Problem

Conventional solution methods for logistics optimization problems often have to be adapted when objectives or restrictions of organizations in logistics environments are changing. In this paper, a new, generic solution approach called optimization network (ON) is developed and applied to a logistics optimization problem, the Location Routing Problem (LRP). With this approach, required flexibility in terms of fast changing data within the advancement of industry 4.0 is addressed. In an ON, existing solution methods are applied to the basic problems of the LRP. A meta solver optimizes the overall result of the network with black box optimization. Based on this, an orchestrator is responsible for the introduction of new optimization runs. The developed approach guarantees that changing external influences only involve the adaption of affected optimization nodes within the ON and not of the whole solution approach. Results are compared with an already existing generic solver and show the potential of the new solution method.

Evolutionary Procedural 2D Map Generation using Novelty Search

This paper presents an evolutionary approach to procedural content generation of 2D maps for computer games. To provide better adaptability to the map designers vision, user preference is incorporated to guide the algorithm. A cooperative method utilizes novelty search as a source of diverse solutions, which are then further optimized by multiple, subsequent genetic algorithms. We compare the results to a second approach based on multi-objective optimization, which takes the two conflicting goals of optimizing towards user preference and finding novel solutions as objective functions to build a Pareto front of maps.

Enterprise Interoperability as Framework for Project Knowledge Management

In this paper we are identifying enterprise interoperability as a possible framework for knowledge management in projects. In particular we are reflecting on research projects, involving small groups of companies and research institutions with heterogeneous backgrounds. The nature of research raises several issues due to the heterogeneous expertise of the project teams. This is in particular true when multiple (semi-formal) models are developed in parallel. In addition to the work evolving over time at the partner’s locations, the organisational contexts moved over time, often invisible to the rest of the teams. Interoperability barriers emerge between organisations and models. Enterprise Interoperability can be used to prepare the project teams, supporting the required work on interfaces between models.

Simplifying Problem Definitions in the HeuristicLab Optimization Environment

Software frameworks for metaheuristic optimization take the burden off researchers and practitioners to start from scratch and implement their own algorithms and problems. One such framework is HeuristicLab. While it allows using existing, already implemented algorithms and problems comfortably and provides an extensive range of tools for analyzing results, it lacks an easy to use programming interface for adding new problems. As implementing new problems is a common task, an improved and simpler problem definition interface has been created. Besides giving an overview of the implementation, we also show examples of problems built using this new interface. Additionally, we compare the new approach to three other metaheuristic frameworks. This is done by analyzing the source code of the OneMax problem implemented in each framework and comparing the resulting lines of code with previous works.

Optimizing Set-Up Times Using the HeuristicLab Optimization Environment

This publication shows the application of set-up time optimization to machinery that requires some components to be preloaded from a component storage to the work zone before jobs can be processed. Component loading and unloading, which is normally done by the machine operators, should be done automatically. The machine has access to a component storage consisting of multiple racks. Components can be moved by using different strategies. These strategies also affect the storage layout over time. Applying simulation-based optimization to the set-up process yields good machine configuration parameters (i.e. initial storage layout, sequence of jobs and used strategies) for a given job sequence. A simulator which models the machinery is used to evaluate different strategies and machine parameters. For all optimization aspects, HeuristicLab is used as the underlying software environment in combination with a new specific problem type that can be solved with evolutionary algorithms such as genetic algorithms or evolution strategies.