Heider, Michael
Refine
Document Type
- Conference Proceeding (15) (remove)
Language
- English (15)
Keywords
- Cartesian Genetic Programming (2)
- Crossover (2)
- CGP (1)
- Evolutionary Algorithm (1)
- Positional Bias (1)
- Recombination (1)
- Reorder (1)
Ant-based metaheuristics have successfully been applied to a variety of different graph-based problems.
However, for Cartesian Genetic Programming (CGP) only the impact of Max-Min Ant Systems has been tested.
In this work, we try to fill this gap by applying four different popular ant-based metaheuristics as the optimizer (and therefore training algorithm) of CGP.
The idea of combining CGP with ant-based metaheuristics is not novel but older works’ experimental design may not meet today’s standard.
To compare these metaheuristics to the Evolution Strategies (ESs) commonly used in CGP, we benchmark against a standard CGP variant that uses a simplistic (1 + 4)-ES, mutation, and no crossover.
Additionally, we include (μ + λ)-ES and (μ, λ)-ES in our experiments.
We analyse the performance on datasets from the symbolic regression, regression, and classification domains.
By tuning and evaluating various configurations, we can not affirm a significant improvement by using ant-based methods with CGP as we encounter premature convergence—even with those ant-based metaheuristics that were originally proposed to overcome such problems.
Despite our results being of negative nature, this work still gives important and interesting insights into the training of CGP models.
The key contributions of our work are thus a more thorough benchmarking of these optimizers than has been done before.
This should clear up doubts about the capabilities of ant-based metaheuristics in CGP.
Furthermore, we include a roadmap on how they can be addressed to solve this complex optimization problem from the model building domain of machine learning.
The Machine Health (MH) sector—which includes, for example, Predictive Maintenance, Prognostics and Health Management, and Condition Monitoring—has the potential to improve efficiency and reduce costs for maintenance and machine operation. This is achieved by data-driven analytics applications, utilising the vast amount of data collected by sensors during machine runtime. While there are numerous possible fields of application, the overall complexity of machines and applications in scientific publications is still low, preventing MH technologies from being implemented in many real-world scenarios. This may be the result of a diffuse understanding of the term complexity in the publications of this field, which results in a lack of focus towards the core problems of real-world MH applications. This article introduces a new way of discerning complexity in data-driven MH applications, enabling an effective discussion and analysis of present and future MH applications. This is achieved by creating a new taxonomy based on observations from relevant literature and substantial domain knowledge. Using this newly introduced taxonomy, we categorise recent applications of MH to demonstrate the usefulness of our approach and illustrate a still-prevalent research gap based on our findings.
The recombination operator plays an important role in many evolutionary algorithms. However, in Cartesian Genetic Programming (CGP), which is part of the aforementioned category, the usefulness of crossover is contested. In this work, we investigate whether CGP’s positional bias actually influences the usefulness of the crossover operator negatively. This bias describes a skewed distribution of CGP’s active and inactive nodes, which might lead to destructive behaviours of standard recombination operators. We try to answer our hypothesis by employing one standard CGP implementation and one without the effects of positional bias. Both versions are combined with one of four standard crossover operators, or with no crossover operator. Additionally, two different selection methods are used to configure a CGP variant. We then analyse their performance and convergence behaviour on eight benchmarks taken from the Boolean and symbolic regression domain. By using Bayesian inference, we are able to rank them, and we found that positional bias does not influence CGP with crossover. Furthermore, we argue that the current research on CGP with standard crossover operators is incomplete, and CGP with recombination might not negatively impact its evolutionary search process. On the contrary, using CGP with crossover improves its performance.
Maintenance of complex machinery is time and resource intensive. Therefore, decreasing maintenance cycles by employing Predictive Maintenance (PdM) is sought after by many manufacturers of machines and can be a valuable selling point. However, currently PdM is a hard to solve problem getting increasingly harder with the complexity of the maintained system. One challenge is to adequately prepare data for model training and analysis. In this paper, we propose the use of expert knowledge–based preprocessing techniques to extend the standard data science–workflow. We define complex multi-purpose machinery as an application domain and test our proposed techniques on real-world data generated by numerous machines deployed in the wild. We find that our techniques enable and enhance model training.
A very common and powerful step in the design process of a new learning algorithm or extensions and improvements of existing algorithms is the benchmarking of models produced by said algorithm. We propose a paradigm shift in the benchmarking of explainable if-then-rule-based models like the ones generated by Learning Classifier Systems or Fuzzy Rule-based Systems. The principled method we suggest is based on synthetic data sets being sampled from randomly generated but known processes that have the same overall structure as the models that are being benchmarked (i. e. each process consists of a set of components each of which corresponds to one if-then rule) which is up-to-date not available among the many synthetic data generators. This approach has several benefits over other benchmarks and we expect that it will lead to fresh insights into how algorithms producing such explainable models work and can be improved. We demonstrate its usage by benchmarking the effects of different rule representations in the XCSF classifier system.
Fast, flexible, and fearless: a rust framework for the modular construction of metaheuristics
(2023)
We present MAHF, a Rust framework for the modular construction and subsequent evaluation of evolutionary algorithms, but also any other metaheuristic framework, including non-population-based and constructive approaches. We achieve high modularity and flexibility by splitting algorithms into components with a uniform interface and allowing communication through a shared blackboard. Nevertheless, MAHF is aimed at being easy to use and adapt to the specific purposes of different practitioners. To this end, this paper focuses on providing a general description of the design of MAHF before illustrating its application with a variety of different use cases, ranging from simple extension of the set of implemented components and the subsequent construction of algorithms not present within the framework to hybridization approaches, which are often difficult to realize in specialized software frameworks. By providing these comprehensive examples, we aim to encourage others to utilize MAHF for their needs, evaluate its effectiveness, and improve upon its application.
Analysing the behaviour of metaheuristics comprehensively and thereby enhancing explainability requires large empirical studies. However, the amount of data gathered in such experiments is often too large to be examined and evaluated visually. This necessitates establishing more efficient analysis procedures, but care has to be taken so that these do not obscure important information. This paper examines the suitability of clustering methods to assist in the characterisation of the behaviour of metaheuristics. The convergence behaviour is used as an example as its empirical analysis often requires looking at convergence curve plots, which is extremely tedious for large algorithmic datasets. We used the well-known K-Means clustering method and examined the results for different cluster sizes. Furthermore, we evaluated the clusters with respect to the characteristics they utilise and compared those with characteristics applied when a researcher inspects convergence curve plots. We found that clustering is a suitable technique to assist in the analysis of convergence behaviour, as the clusters strongly correspond to the grouping that would be done by a researcher, though the procedure still requires background knowledge to determine an adequate number of clusters. Overall, this enables us to inspect only few curves per cluster instead of all individual curves.
Unlike in traditional Genetic Programming, Cartesian Genetic Programming (CGP) does not commonly feature a recombination/crossover operator, although recombination plays an important role in other evolutionary techniques, including Genetic Programming from which CGP originates. Instead, CGP mainly depends on mutation and selection operators in their evolutionary search. To this day, it is still unclear as to why CGP’s performance does not generally improve with the addition of crossover. In this work, we argue that CGP’s positional bias might be a reason for this phenomenon. This bias describes a skewed distribution of active and inactive nodes, which might lead to destructive behaviour of standard recombination operators. We provide a first assessment with preliminary results. No final conclusion to this hypothesis can be drawn yet, as more thorough evaluations must be done first. However, our first results show promising trends and may lay the foundationf or future work.
To fill the increasing demand for explanations of decisions made by automated prediction systems, machine learning (ML) techniques that produce inherently transparent models are directly suited. Learning Classifier Systems (LCSs), a family of rule-based learners, produce transparent models by design. However, the usefulness of such models, both for predictions and analyses, heavily depends on the placement and selection of rules (combined constituting the ML task of model selection). In this paper, we investigate a variety of techniques to efficiently place good rules within the search space based on their local prediction errors as well as their generality. This investigation is done within a specific LCS, named SupRB, where the placement of rules and the selection of good subsets of rules are strictly separated in contrast to other LCSs where these tasks sometimes blend. We compare a Random Search, (1,λ)-ES and three Novelty Search variants. We find that there is a definitive need to guide the search based on some sensible criteria, i.e. error and generality, rather than just placing rules randomly and selecting better performing ones but also find that Novelty Search variants do not beat the easier to understand (1,λ)-ES.
To increase the utility of semantic industrial information models we propose a methodology to incorporate extracted operator knowledge, which we assume to be present in the form of rules, in knowledge graphs. To this end, we present multiple modelling patterns that can be combined depending on the required complexity. Aiming to combine information models with learning systems we contemplate desired behaviours of embeddings from a predictive quality perspective and provide a suited embedding methodology. This methodology is evaluated on a real world dataset of a fused deposition modelling process.
While knowledge graphs and their embedding into low dimensional vectors are established fields of research, they mostly cover factual knowledge. However, to improve downstream models, e. g. for predictive quality in real-world industrial use cases, embeddings of procedural knowledge, available in the form of rules, could be utilized. As such, we investigate which properties of embedding algorithms could prove beneficial in this scenario and evaluate which established embedding methodologies are suited to form the basis of sum-based embeddings of different representations of procedural knowledge.
Just-in-time supply chains have become increasingly popular in past decades. However, these are particularly vulnerable when logistic routes are blocked, manufacturing capacities are limited or customs are under strain, as has been seen in the last few years. The principle of just-in-time delivery requires a coordinated production and material flow along the entire supply chain. Challenges in the supply chain can lead to various disruptions, so that certain manufacturing jobs must be changed, postponed or cancelled, which will then impact supply down the line up to the consumer. Nowadays, many planning and control processes in the event of a disturbance are based on the procedural knowledge of employees and undertaken manually by those. The procedures to mitigate the negative effects of disturbances are often quite complex and time-critical, making disturbance management highly challenging. In this paper, we introduce a real-world use case where we automate the currently manual reschedule of a production plan containing unavailable jobs. First, we analyse existing literature regarding the classification of disturbances encountered in similar use cases. We show how we automate existing manual disturbance management and argue that employing stochastic optimization allows us to not only promote future jobs but to on-the-fly create entirely new plans that are optimized regarding throughput, energy consumption, material waste and operator productivity. Building on this routine, we propose to create a Bayesian estimator to determine the probabilities of delivery times whose predictions we can then reintegrate into our optimizer to create less fragile schedules. Overall, the goals of this approach are to increase robustness in production planning and control.
