This paper introduces a novel Q-learning framework to address instabilities in offline reinforcement learning with continuous state spaces.We identify the recurring collapse of Q-value targets as core challenge and propose a stabilization technique that replaces the iteration-wise targets with their point-wise maximum across iterations.This approach enforces convergence and fully mitigates recursive errors.We show that a performance metric linking Q-values to policy performance is directly available.Our findings represent a first step toward stabilizing Q-learning in challenging settings and highlight the potential of model-based approaches.

Machine learning is rapidly transforming industrial landscapes, yet it faces significant hurdles related to efficiency and resilience.This paper discusses industrial challenges and provides a structured overview of current approaches, encompassing data-centric methodologies, efficient training for reliable solutions, hardware-optimized deployment, and the emerging role of foundation models. * This work was

Possibilistic c-means (PCM) clustering began in 1993, and has been used since then in many applications. In this article we discuss the geometric foundations of PCM and introduce a new hard possibilistic c-means (HPCM) clustering algorithm. We use limit theory to prove that the extended set of possibilistic c-partitions is the unit hypercube in R c n ; and that its vertices are exactly the hard possibilistic c-partitions on n objects defined herein. This enables completion of the geometric description of the domain of possibilistic clustering algorithms. We give examples that compare the results of clustering with Hard c-means (HCM) to HPCM on three small synthetic data sets. Our proof-of-concept examples show that the new algorithm performs as expected, and provides much more realistic interpretation of clusters than HCM when the data contain bridge points or noise.

In this paper, we investigate offline reinforcement learning (RL) with the goal of training a single robust policy that generalizes effectively across environments with unseen dynamics.We propose a novel approach, Trajectory Encoding Augmentation (TEA), which extends the state space by integrating latent representations of environmental dynamics obtained from sequence encoders, such as autoencoders.Our findings show that incorporating these encodings with TEA improves the transferability of a single policy to novel environments with new dynamics, surpassing methods that rely solely on unmodified states.These results indicate that TEA captures critical, environment-specific characteristics, enabling RL agents to generalize effectively across dynamic conditions.

This paper investigates the instability of Q-learning in continuous environments, a challenge frequently encountered by practitioners.Traditionally, this instability is attributed to bootstrapping and regression model errors.Using a representative reinforcement learning benchmark, we systematically examine the effects of bootstrapping and model inaccuracies by incrementally eliminating these potential error sources.Our findings reveal that even in relatively simple benchmarks, the fundamental task of Q-learning -iteratively learning a Q-function from policy-specific target values -can be inherently ill-posed and prone to failure.These insights cast doubt on the reliability of Q-learning as a universal solution for reinforcement learning problems.

Interest in solving table interpretation tasks has grown over the years, yet it still relies on existing datasets that may be overly simplified. This is potentially reducing the effectiveness of the dataset for thorough evaluation and failing to accurately represent tables as they appear in the real-world. To enrich the existing benchmark datasets, we extract and annotate a new, more challenging dataset. The proposed Wiki-TabNER dataset features complex tables containing several entities per cell, with named entities labeled using DBpedia classes. This dataset is specifically designed to address named entity recognition (NER) task within tables, but it can also be used as a more challenging dataset for evaluating the entity linking task. In this paper we describe the distinguishing features of the Wiki-TabNER dataset and the labeling process. In addition, we propose a prompting framework for evaluating the new large language models on the within tables NER task. Finally, we perform qualitative analysis to gain insights into the challenges encountered by the models and to understand the limitations of the proposed~dataset.

In subtractive manufacturing, predicting the lifespan of tools offers significant advantages. However, data-driven methods for such predictions often fail to incorporate domain knowledge effectively. This paper presents a hybrid approach that combines machine learning with domain expertise. Firstly, we introduce a novel method to augment sparsely labeled datasets. Secondly, we propose a new loss function that integrates domain knowledge into a machine learning model. This approach enhances the accuracy and reliability of tool wear predictions, ultimately improving the efficiency of subtractive manufacturing processes. We evaluate the proposed methods on the NUAA Ideahouse Dataset. We achieve a R2score of up to 0.99 on unseen data.

Large Language Models (LLMs) have provided a new pathway for Named Entity Recognition (NER) tasks. Compared with fine-tuning, LLM-powered prompting methods avoid the need for training, conserve substantial computational resources, and rely on minimal annotated data. Previous studies have achieved comparable performance to fully supervised BERT-based fine-tuning approaches on general NER benchmarks. However, none of the previous approaches has investigated the efficiency of LLM-based few-shot learning in domain-specific scenarios. To address this gap, we introduce FsPONER, a novel approach for optimizing few-shot prompts, and evaluate its performance on domain-specific NER datasets, with a focus on industrial manufacturing and maintenance, while using multiple LLMs – GPT-4-32K, GPT-3.5-Turbo, LLaMA 2-chat, and Vicuna. FsPONER consists of three few-shot selection methods based on random sampling, TF-IDF vectors, and a combination of both. We compare these methods with a general-purpose GPT-NER method as the number of few-shot examples increases and evaluate their optimal NER performance against fine-tuned BERT and LLaMA 2-chat. In the considered real-world scenarios with data scarcity, FsPONER with TF-IDF surpasses fine-tuned models by approximately 10% in F1 score.

Recent advances in Tiny Machine Learning (TinyML) empower low-footprint embedded devices for real-time on-device Machine Learning (ML). While many acknowledge the potential benefits of TinyML, its practical implementation presents unique challenges. This study aims to bridge the gap between prototyping single TinyML models and developing reliable TinyML systems in production: (1) Embedded devices operate in dynamically changing conditions. Existing TinyML solutions primarily focus on inference, with models trained offline on powerful machines and deployed as static objects. However, static models may underperform in the real world due to evolving input data distributions. We propose online learning to enable training on constrained devices, adapting local models toward the latest field conditions. (2) Nevertheless, current on-device learning methods struggle with heterogeneous deployment conditions and the scarcity of labeled data when applied across numerous devices. We introduce federated meta-learning incorporating online learning to enhance model generalization, facilitating rapid learning. This approach ensures optimal performance among distributed devices by knowledge sharing. (3) Moreover, TinyML’s pivotal advantage is widespread adoption. Embedded devices and TinyML models prioritize extreme efficiency, leading to diverse characteristics ranging from memory and sensors to model architectures. Given their diversity and non-standardized representations, managing these resources becomes challenging as TinyML systems scale up. We present semantic management for the joint management of models and devices at scale. We demonstrate our methods through a basic regression example and then assess them in three real-world TinyML applications: handwritten character image classification, keyword audio classification, and smart building presence detection. The results confirm the effectiveness of our approaches from various perspectives, such as accuracy improvement, resource savings, and engineering effort reduction.

Recently, offline RL algorithms have been proposed that remain adaptive at runtime.For example, the LION algorithm [1] provides the user with an interface to set the trade-off between behavior cloning and optimality w.r.t. the estimated return at runtime.Experts can then use this interface to adapt the policy behavior according to their preferences and find a good trade-off between conservatism and performance optimization.Since expert time is precious, we extend the methodology with an autopilot that automatically finds the best parameterization of the trade-off, yielding a new algorithm which we term AutoLION.

Tiny machine learning (TinyML) is a rapidly growing field aiming to democratize machine learning (ML) for resource-constrained microcontrollers (MCUs). Given the pervasiveness of these tiny devices, it is inherent to ask whether TinyML applications can benefit from aggregating their knowledge. Federated learning (FL) enables decentralized agents to jointly learn a global model without sharing sensitive local data. However, a common global model may not work for all devices due to the complexity of the actual deployment environment and the heterogeneity of the data available on each device. In addition, the deployment of TinyML hardware has significant computational and communication constraints, which traditional ML fails to address. Considering these challenges, we propose TinyReptile, a simple but efficient algorithm inspired by meta-learning and online learning, to collaboratively learn a solid initialization for a neural network (NN) across tiny devices that can be quickly adapted to a new device with respect to its data. We demonstrate TinyReptile on Raspberry Pi 4 and Cortex-M4 MCU with only 256-KB RAM. The evaluations on various TinyML use cases confirm a resource reduction and training time saving by at least two factors compared with baseline algorithms with comparable performance.

Decision making aims to select good decision options, based on ratings of utility. It is often difficult for experts to provide exact numerical ratings for decision options, so they rather prefer interval valued ratings. Interval type–2 fuzzy decision making is based on such interval valued ratings. The decision process has to balance between best case and worst case ratings, depending on the acceptable level of risk. The PIU method finds decisions by Pareto optimization of best case and worst case ratings. This paper introduces an extension of PIU to labeled objects: LPIU. Experiments with the car preference data set show that LPIU not only allows to find good decision options but also to explain how the object labels affect the decision process, an important step towards explainable AI.

The Internet of Things (IoT) is revolutionizing the industry. Powered by pervasive embedded devices, the Industrial IoT (IIoT) provides a unique solution for retrieving and analyzing data near the source in real-time. Many emerging techniques, such as Tiny Machine Learning (TinyML) and Complex Event Processing (CEP) , are actively being developed to support decision making at the edge, shifting the paradigm from centralized processing to distributed computing. However, distributed computing presents management challenges, as IoT devices are diverse and constrained, and their number is growing exponentially. The situation is even more challenging when various on-device applications (so-called artifacts) are deployed across decentralized IoT networks. Questions to be addressed include how to discover an appropriate function, whether that function can be executed on a certain device, and how to orchestrate a cross-platform service. To tackle these challenges, we propose an approach for the scalable management of on-device applications among distributed IoT devices. By leveraging the W3C Web of Things (WoT) , the capabilities of each IoT device, or more precisely, its interaction patterns, can be semantically expressed in a Thing Description (TD) . In addition, we introduce semantic modeling of on-device applications to supplement an TD with additional information regarding applications on the device. Specifically, we demonstrate two examples of semantic modeling: neural networks (NN) and CEP rules. The ontologies are evaluated by answering a set of competency questions. By hosting the enriched semantic knowledge of the entire IoT system in a Knowledge Graph (KG) , we can discover and interoperate edge devices and artifacts across the decentralized network. This can reduce fragmentation and increase the reusability of IoT components. We demonstrate the feasibility of our concept on an industrial workstation consisting of a conveyor belt and several IoT devices. Finally, the requirements for constructing an IoT semantic management system are discussed.

Privacy issues prevent data owner from improving Machine Learning (ML) performance as it makes external collaborations binding.To allow data sharing without confidentiality concerns, we propose in this work methods to generate time series in a privacy preserving manner.We combine the existing Generative Adversarial Networks (GAN) models for time series namely TimeGAN [1], ClaRe-GAN [2] and C-RNN-GAN [3] with differential privacy.This is achieved by changing their original discriminator with a private discriminator that relies on the differentially private stochastic gradient method (DPSGD) [4].Our experiments show that the developed methods -in particular TimeGAN and ClaRe-GANoutperform the existing and unique differentially private model for time series of RCGAN [5] in terms of privacy and accuracy.

In offline reinforcement learning, a policy needs to be learned from a single pre-collected dataset.Typically, policies are thus regularized during training to behave similarly to the data generating policy, by adding a penalty based on a divergence between action distributions of generating and trained policy.We propose a new algorithm, which constrains the policy directly in its weight space instead, and demonstrate its effectiveness in experiments. *The project this paper is based on was supported with funds from the German Federal Ministry of

In the last few years, several works have been proposed on Generative Adversarial Networks (GAN). At the same time, there is a lack of investigation on their evaluation and the few proposed evaluation methods have not yet been rigorously studied. By way of example, the metrics to evaluate Generative Adversarial Networks (GAN) have been exclusively developed and tested for image data, but models operating on time series data have not been studied at all. In fact, it is still unclear what are the advantages and disadvantages of each approach and what is the difference between them in terms of performance i.e. which metric can easily detect common GAN problems such as mode collapse or mode dropping. Different tests have been introduced by [6] to evaluate GAN metrics for images. Inspired by this work, we extensively study the numerous evaluation metrics proposed in the literature that are designed for time series data and compare them to each other in a structured way. To the best of our knowledge, this represents the first work that studies the existing evaluation metrics of GAN for time series data and evaluates their performance against different evaluation criteria. Moreover, we introduce MiVo, a new evaluation metric that computes the similarity between a set of real and a set of generated data and finds for each real times series a synthetic one and for each synthetic time series a real one. We show that this bidirectional check will allow to easily detect different training problems such as the ones mentioned above. At the same time, this method is computationally much more efficient as it doesn't involve any machine learning model and hence no training is needed.

Tiny machine learning (TinyML) is a fast-growing research area committed to democratizing deep learning for all-pervasive microcontrollers (MCUs). Challenged by the constraints on power, memory, and computation, TinyML has achieved significant advancement in the last few years. However, the current TinyML solutions are based on batch/offline setting and support only the neural network's inference on MCUs. The neural network is first trained using a large amount of pre-collected data on a powerful machine and then flashed to MCUs. This results in a static model, hard to adapt to new data, and impossible to adjust for different scenarios, which impedes the flexibility of the Internet of Things (IoT). To address these problems, we propose a novel system called TinyOL (TinyML with Online-Learning), which enables incremental on-device training on streaming data. TinyOL is based on the concept of online learning and is suitable for constrained IoT devices. We experiment TinyOL under supervised and unsupervised setups using an autoencoder neural network. Finally, we report the performance of the proposed solution and show its effectiveness and feasibility.

The increasing importance of resource-efficient production entails that manufacturing companies have to create a more dynamic production environment, with flexible manufacturing machines and processes. To fully utilize this potential of dynamic manufacturing through automatic production planning, formal skill descriptions of the machines are essential. However, generating those skill descriptions in a manual fashion is labor-intensive and requires extensive domain-knowledge. In this contribution an ontology-based semi-automatic skill description system that utilizes production logs and industrial ontologies through inductive logic programming is introduced and benefits and drawbacks of the proposed solution are evaluated.

Many decision making processes are based on choosing options with maximum utility. Often utility assessments are associated with uncertainty, which may be mathematically modeled by intervals of utilities. Intervals of utilities may be mapped to single utility values by so-called type reduction methods which have been originally developed in the context of interval type-2 defuzzification: the method by Nie and Tan (NT), consistent linear type reduction (CLTR), consistent quadratic type reduction (CQTR), and the uncertainty weight method (UW). This paper considers the problem of comparing pairs of utility intervals using type reduction methods. Three different possible relations between pairs of intervals (disjoint, overlapping, and inclusive) are distinguished in an extensive experimental study, which yields recommendations for the choice of type reduction methods with respect to the level of risk that the decision maker is willing to take. If the focus is on mean utility, then we recommend the Nie-Tan method. For more cautious decision making, when very low utilities should be avoided, we recommend consistent linear type reduction with a high value of the cautiousness parameter or consistent quadratic type reduction. For more risky decision making with a strong focus on very high utilities we recommend consistent linear type reduction with a low value of the cautiousness parameter.

Lifelong learning has recently attracted attention in building machine learning systems that continually accumulate and transfer knowledge to help future learning. Unsupervised topic modeling has been popularly used to discover topics from document collections. However, the application of topic modeling is challenging due to data sparsity, e.g., in a small collection of (short) documents and thus, generate incoherent topics and sub-optimal document representations. To address the problem, we propose a lifelong learning framework for neural topic modeling that can continuously process streams of document collections, accumulate topics and guide future topic modeling tasks by knowledge transfer from several sources to better deal with the sparse data. In the lifelong process, we particularly investigate jointly: (1) sharing generative homologies (latent topics) over lifetime to transfer prior knowledge, and (2) minimizing catastrophic forgetting to retain the past learning via novel selective data augmentation, co-training and topic regularization approaches. Given a stream of document collections, we apply the proposed Lifelong Neural Topic Modeling (LNTM) framework in modeling three sparse document collections as future tasks and demonstrate improved performance quantified by perplexity, topic coherence and information retrieval task. Code: https://github.com/pgcool/Lifelong-Neural-Topic-Modeling

Many manufacturing processes ask for greater flexibility and transparency, which represent key features for constructing robust industrial cyber-physical systems. In the Industry 4.0 scenario, where factories are equipped with 'smart' machines, it is crucial to have robust AI methods that allow to completely automate assembly by quickly adapting to various production requests. This is complemented by making the system transparent, which gives understanding how the manufacturing process is carried out and helps build trust in such systems. To this end, we exploit semantic technologies to represent machines' capabilities and match them to production requests, to address the question of flexibility, as well as AI planning techniques to generate production sequences. We utilize OWL justifications, as a well established explanations technique in OWL ontologies, to explain the matching of machines' skills to product requests and the generated production sequences. We illustrate results from our methods for an experimental production facility.

In the last few years, a huge progress has been made to achieve image-to-image translation by mapping images from a source domain to a target domain. We exploit the recent progress made in this field to tackle another issue, namely time series translation. This work targets time series translation i.e. maps time series data from a source domain to a target domain. We present our new algorithm DR-TiST, a modified version of DRIT [1], that enables time series translation. We apply DR-TiST to a real word use case where we transfer the time series behavior of a ventilation system to the environmental conditions of a different ventilation system and introduce new evaluation metrics to evaluate its performance. The performance of DR-TiST is compared to CycleGAN-VC [14], a special form of an image-to-image translation algorithm used for voice conversion. We demonstrate that the time series generated by DR-TiST are more realistic than the ones generated by CycleGAN-VC.

Past work in relation extraction mostly focuses on binary relation between entity pairs within single sentence. Recently, the NLP community has gained interest in relation extraction in entity pairs spanning multiple sentences. In this paper, we propose a novel architecture for this task: inter-sentential dependency-based neural networks (iDepNN). iDepNN models the shortest and augmented dependency paths via recurrent and recursive neural networks to extract relationships within (intra-) and across (inter-) sentence boundaries. Compared to SVM and neural network baselines, iDepNN is more robust to false positives in relationships spanning sentences. We evaluate our models on four datasets from newswire (MUC6) and medical (BioNLP shared task) domains that achieve state-of-the-art performance and show a better balance in precision and recall for inter-sentential relationships. We perform better than 11 teams participating in the BioNLP shared task 2016 and achieve a gain of 5.2% (0.587 vs 0.558) in F1 over the winning team. We also release the crosssentence annotations for MUC6.

We propose the application of interval type-2 fuzzy decision making (IT2FDM) to dynamic scheduling of deliveries in a just-in-time logistic process. Delivery decisions are based on order priorities computed from the expected decrease of customer satisfaction for each order. We compare IT2FDM with first in first out (FIFO), earliest due date first (EDDF), and (type-1) fuzzy decision making (FDM). In a simulation of a real world process for a duration of 600 days IT2FDM in comparison with the three other methods yields the highest just-in-time delivery rate, the highest average customer satisfaction, and the highest percentage of very satisfied customers. The increasing percentage of very satisfied customers at the same time leads to a slightly increasing percentage of not satisfied customers, but the trade- off between these percentages can be balanced by an appropriate choice of the risk parameter.

The Possibilistic C-Means (PCM) was developed as an extension of the Fuzzy C-Means (FCM) by abandoning the membership sum-to-one constraint. In PCM, each cluster is independent of the other clusters, and can be processed separately. Thus, the Sequential Possibilistic One-Means (SP1M) was proposed to find clusters sequentially by running P1M C times. One critical problem in both PCM and SP1M is how to determine the parameter η. The Sequential Possibilistic One Means with Adaptive Eta (SP1M-AE) was developed to allow η to change during iterations. In this paper, we introduce a new dynamic adaption mechanism for the parameter η in each cluster and apply it into SP1M. The resultant algorithm, called the Sequential Possibilistic One-Means with Dynamic Eta (SP1M-DE) is shown to provide superior performance over PCM, SP1M, and SP1M-AE in determining correct clustering results.

We propose a novel Bayesian approach to modelling nonlinear alignments of time series based on latent shared information. We apply the method to the real-world problem of finding common structure in the sensor data of wind turbines introduced by the underlying latent and turbulent wind field. The proposed model allows for both arbitrary alignments of the inputs and non-parametric output warpings to transform the observations. This gives rise to multiple deep Gaussian process models connected via latent generating processes. We present an efficient variational approximation based on nested variational compression and show how the model can be used to extract shared information between dependent time series, recovering an interpretable functional decomposition of the learning problem. We show results for an artificial data set and real-world data of two wind turbines.

Rule-based diagnostics of power generating equipment is an important task in industry. In this demo we present how semantic technologies can enhance diagnostics. In particular, we present our semantic rule language sigRL that is inspired by the real diagnostic languages in Siemens. SigRL allows to write compact yet powerful diagnostic programs by relying on a high level data independent vocabulary, diagnostic ontologies, and queries over these ontologies. We present our diagnostic system SemDia. The attendees will be able to write diagnostic programs in SemDia using sigRL over 50 Siemens turbines. We also present how such programs can be automatically verified for redundancy and inconsistency. Moreover, the attendees will see the provenance service that SemDia provides to trace the origin of diagnostic results.

Fuzzy c-means (FCM) clustering is known to be sensitive to outliers and noise. Possibilistic c-means (PCM) has been reported to be more robust against outliers and noise but may yield coincident clusters. We introduce a variant of PCM called sequential possibilistic one-means (SP1M) that finds clusters sequentially, takes into account the previously found clusters for initialization, and discards coincident clusters. Experiments with the well-known BIRCH benchmark data set and two variants of BIRCH indicate that SP1M is able to find a significantly larger percentage of the clusters contained in the data, with about twice as many cluster update steps, but significantly faster than FCM and PCM.

The Particle Swarm Optimization Policy (PSO-P) has been recently introduced and proven to produce remarkable results on interacting with academic reinforcement learning benchmarks in an off-policy, batch-based setting. To further investigate the properties and feasibility on real-world applications, this paper investigates PSO-P on the so-called Industrial Benchmark (IB), a novel reinforcement learning (RL) benchmark that aims at being realistic by including a variety of aspects found in industrial applications, such as continuous state and action spaces, a high dimensional, partially observable state space, delayed effects, and complex stochasticity. The experimental results of PSO-P on IB are compared to results of closed-form control policies derived from the model-based Recurrent Control Neural Network (RCNN) and the model-free Neural Fitted Q-Iteration (NFQ). Experiments show that PSO-P is not only of interest for academic benchmarks, but also for real-world industrial applications, since it also yielded the best performing policy in our IB setting. Compared to other well established RL techniques, PSO-P produced outstanding results in performance and robustness, requiring only a relatively low amount of effort in finding adequate parameters or making complex design decisions.

In industrial power generation plants, subsystem monitoring and analytics play a vital role in quantifying the knowledge about different factors that impact their overall performance. Multi-dimensional performance metrics, e.g. thermal efficiency, in-service time, mean-time-to-failure etc., are calculated that may have different data constraints, modelling techniques, and execution frameworks. Automating these calculations and combining multiple metrics to form a single performance index (e.g. reliability) is a challenging task as it requires considerable domain-specific expertise and consolidation of performance-related data and its underlying models. In this paper, we propose to use ontologies to assist domain analyst to first, capture appropriate semantic data of an individual performance metric, and later to provide means to integrate and execute multiple metrics to accurately reflect the overall performance of a plant. We present our prototypical implementation, its evaluation; furthermore, we discuss an ontology model that currently describes three distinct analytical models and its related data based on the case study of Siemens gas turbines. We also demonstrate how ontologies can support to infer the appropriate aggregation method in calculating composite indices.

Our goal is to automatically generate membership functions for linguistic terms such as newborn, baby, child, teen, and adult. Many approaches have been proposed to generate membership functions from data, for example using fuzzy clustering or neural networks, but these generate abstract membership functions with no semantic meaning. Our approach uses word vectors that are extracted from large text corpora, so that semantically similar words have similar word vectors. We use projections of word vectors to (i) generate the peaks of membership functions over one-dimensional domains and use similarities between pairs of word vectors to (ii) compute selected membership values. We present four different alternatives to construct membership functions from (i) and (ii), which produce (a) singleton functions, (b) piecewise linear functions, (c) triangular functions, and (d) normalized triangular functions. In the experimental part we present examples from four different domains where our method successfully generates linguistically meaningful membership functions from word vectors of the 50-dimensional glove.6B.50d data set, and we also present an example that illustrates the limitations of the word vector based approach, when the considered linguistic terms come from different domains or contain ambiguity.

The C index is an internal cluster validity index that was introduced in 1970 as a way to define and identify a “best” crisp partition on n objects represented by either unlabeled feature vectors or dissimilarity matrix data. This index is often one of the better performers among the plethora of internal indices available for this task. This paper develops a soft generalization of the C index that can be used to evaluate sets of candidate partitions found by either fuzzy or probabilistic clustering algorithms. We define four generalizations based on relational transformations of the soft partition and, then, compare their performance to eight other popular internal fuzzy cluster indices using two methods of comparison (internal “best-c ” and internal/external (I/E) “best match”), six synthetic datasets, and six real-world labeled datasets. Our main conclusion is that the sum-min generalization is the second best performer in the best-c tests and the best performer in the I/E tests on small data.

Interval type-2 defuzzification maps an interval type-2 fuzzy set to a crisp number. We show that the semantic meaning of the interval type-2 fuzzy set (the associated opportunity or risk) has to be considered in the choice of an appropriate interval type-2 defuzzification method. Motivated by a list of “axioms” for type-1 defuzzification we introduce twelve mathematical properties for interval type-2 defuzzification that serve as a theoretical framework to assess different interval type-2 defuzzification methods. We show that the well-known Karnik-Mendel algorithm violates at least four of these twelve properties.

We introduce a new method for multidimensional scaling in dissimilarity data that is based on preservation of metric topology between the original and derived data sets. The model seeks neighbors in the derived data that have the same ranks as in the input data. The algorithm we use to optimize the model is a modification of particle swarm optimization called multiswarming. We compare the new method to three well known approaches: Principal component analysis, Sammon's method, and (Kruskal's) metric MDS. Our method produces feature vector realizations that compare favorably with the other approaches on three real relational data sets.

We introduce a new method for feature extraction from object data that is based on the idea of preserving metric topology between the original and derived data sets. Specifically, our method attempts to produce neighbors in the derived data that have the same ranks as in the input data. The algorithm we propose is a novel modification of particle swarm optimization that involves multiswarms. We compare our model and algorithm to feature extraction using Sammon's method and principal components analysis on 19 data sets: 17 are created by making draws from p-variate Gaussian distributions. We also use two real world data sets - the Glass and Lung Cancer data available at the UCI ML website. We find that the new method compares well with Sammon's method, and seems to be superior to features derived with principal components analysis.

A popular clustering model is hard c-means (HCM). For many data sets the HCM objective function has local extrema, so HCM optimization often yields suboptimal clusterings. The effect of local extrema can be reduced by fuzzification, leading to the well-known fuzzy c-means (FCM) model with the fuzziness parameter m >; 1. In this paper we use FCM to optimize the HCM model, even though we actually optimize a different objective function. This work is motivated by a popular approach to avoid local extrema in HCM which approximates the minimum operator in HCM by the harmonic means, leading to c-harmonic means (CHM), which was recently shown to be equivalent to FCM for m = 2. Generalizing the harmonic means in CHM to generalized means yields a clustering model that we call c-generalized means (CGM), which is equivalent to FCM for arbitrary m >; 1. Numerical experiments with the BIRCH and Lena data sets show that FCM/CGM (with optimal m) often yields significantly better HCM clusterings than HCM itself or CHM.

The prerequisite for practicable program analysis is the identification of the individual procedures, which correspond to individual stack frames. We present how machine learning techniques can be used in the setting of program analysis in order to find these stack frames. This combination of machine learning and abstract interpretation-based analysis provides the first fully automatic analysis framework for executables. Our approach can also be applied to identify library functions or malicious behaviour in a given piece of assembly.

A popular algorithm for finding clusters in unlabeled data optimizes the k-means clustering model. This algorithm converges quickly but is sensitive to initialization. Two ways to overcome this drawback are fuzzification and harmonic means. We show that k-harmonic means is a special case of reformulated fuzzy k-means. The main focus of this paper is on partially supervised clustering. Partially supervised clustering finds clusters in data sets that contain both unlabeled and labeled data. We review partially supervised k-means, partially supervised fuzzy k-means, and introduce a partially supervised extension of k-harmonic means. Experiments with four benchmark data sets indicate that partially supervised k-harmonic means inherits the advantages of its completely unsupervised variant: It is significantly less sensitive to initialization than partially supervised k-means.

We propose a novel approach to relational clustering: Given a matrix of pairwise similarity values between objects our algorithm computes a partition of the objects such that similar objects belong to the same cluster and dissimilar objects belong to different clusters. The proposed approach is based on the assumption that the given similarities are products of cluster membership variables. It is based on eigen vector decomposition and minimizes the squared error between the similarities and the products of membership vectors in an efficient, non-iterative way with guaranteed global optimality. In experiments with real world data we show superior performance to conventional iterative clustering approaches.

Fuzzy c-neural network models (FCNNM) combine clustering techniques with advanced neural networks for time series modeling in order to make predictions for a possibly large set of time series using only a small number of models. Given a set of time series, FCNNM finds a partition matrix that quantifies to which degree each time series is associated with each prediction model, as well as the parameters of the neural network models for each cluster. FCNNM allows to automatically identify groups of time series with similar dynamics. This results in higher data efficiency, being of particular interest in cases of poor data availability. We illustrate the application of FCNNM to cash withdrawal series as part of an effective cash management.

Bayesian networks (BNs) are knowledge representation tools capable of representing dependence or independence relationships among random variables. Learning the structure of BNs from datasets has received increasing attention in the last two decades, due to the BNs' capacity of providing good inference models and discovering the structure of complex domains. One approach for BNs' structure learning from data is to define a scoring metric that evaluates the quality of the candidate networks, given a dataset, and then apply an optimization procedure to explore the set of candidate networks. Among the most frequently used optimization methods for BN score-based learning is greedy hill climbing (GHC) search. This paper proposes a new local discovery ant colony optimization (ACO) algorithm and a hybrid algorithm max-min ant colony optimization (MMACO), based on the local discovery algorithm max-min parents and children (MMPC) and ACO to learn the structure of a BN. In MMACO, MMPC is used to construct the skeleton of the BN and ACO is used to orientate the skeleton edges, thus returning the final structure. The algorithms are applied to several sets of benchmark networks and are shown to outperform the GHC and simulated annealing algorithms.

Bayesian networks (BNs) are knowledge representation tools capable of representing dependence or independence relationships among random variables that compose a problem domain. Bayesian networks learned from data sets are receiving increasing attention within the community of researchers of uncertainty in artificial intelligence, due to their capacity to provide good inference models and to discover the structure of complex domains. One approach to learning BNs from data is to use a scoring metric to evaluate the fitness of any given candidate network for the database, and apply an optimization procedure to explore the set of candidate networks. Among the most frequently used optimization methods for this purpose is greedy search, either deterministic or stochastic. This article proposes a hybrid Bayesian network learning algorithm MMACO, based on the local discovery algorithm max-min parents and children (MMPC) and ant colony optimization (ACO). MMPC is used to construct the skeleton of the Bayesian network and then ACO is used to orientate its edges, thus returning the final structure. We apply MMACO (max-min ACO) to several sets of benchmark networks and show that it outperforms greedy search (GS) and simulated annealing (SA) algorithms.

Fuzzy c-auto regression models (FCARM) combine clustering with time series prediction. Given a set of time series, FCARM finds clusters of time series with similar dynamics. More specifically, FCARM finds a partition matrix that quantifies to which degree each time series is associated with each prediction model, and the parameters of the (linear) auto regression models for each cluster. FCARM can thus be used for two different purposes: (i) the automatic identification of clusters of time series with similar dynamics and (ii) the forecast of a large number of time series using only a small number of generic forecast models, leading to higher data efficiency and lower model validation and maintenance effort. We illustrate the application of FCARM to sales forecasts for products that can be clustered into groups with similar sales dynamics.

One of the most important techniques in data preprocessing for data mining is feature selection. Real-world data analysis, data mining, classification and modeling problems usually involve a large number of candidate inputs or features. Less relevant or highly correlated features decrease, in general, the classification accuracy, and enlarge the complexity of the classifier. The goal is to find a reduced set of features that reveals the best classification accuracy for a fuzzy classifier. This paper proposes an ant colony optimization (ACO) algorithm for feature selection, which minimizes two objectives: the number of features and the error classification. Two pheromone matrices and two different heuristics are used for each objective. The performance of the method is compared to other features selection methods, revealing higher performance.

This paper deals with clustering by optimizing the c-means clustering model. For some data sets this clustering model possesses many local optima, so conventional alternating optimization (AO) will produce bad results. For obtaining good clustering results, the minimization procedure has to be kept from being trapped in these local optima, for example, by stochastic optimization approaches. Recently, we showed that ant colony optimization (ACO) can be effectively applied to the c-means clustering model. In this paper, we introduce a wasp swarm optimization (WSO) algorithm to optimize the c-means clustering model. In experiments with four benchmark data sets, the new WSO clustering algorithm is compared with AO and ACO. For data sets leading to c-means models without local optima, both WSO and AO perform better and faster than ACO. For data sets leading to multiple local optima, WSO clearly outperforms both AO and ACO. © 2008 Wiley Periodicals, Inc.

Clustering is often done by minimizing an objective function of a clustering model. Several runs with different initializations or parameters yield multiple solutions. The best of these solutions is often selected by cluster validity measures. We analyze cluster objective and validity functions and show that they can be contradictory and therefore should be considered jointly in an integrated clustering approach. For this reason we define aPareto fuzzy c-means clusteringmodel that produces the Pareto optimal set of both objective and validity functions. In our experiments with the single outlier and the lung cancer data sets Pareto clustering considerably outperforms conventional clustering.

This paper deals with fuzzy clustering by minimizing the fuzzy c-means (FCM) model. We introduce two new methods for minimizing the two reformulated versions of the FCM objective function by particle swarm optimization (PSO). In PSO-V each particle represents a component of a cluster center. In PSO-U each particle represents an unsealed and unnormalized membership value. PSO-V and PSO-U are compared with alternating optimization (AO) and with ant colony optimization (ACO) on two benchmark data sets: the single outlier and the lung cancer data sets. The stochastic methods ACO, PSO-V, and PSO-U are slower than AO, but in each experiment one of the two PSO variants significantly outperforms the other algorithms.

This paper deals with clustering relational data that can be (at least approximately) represented by object data with ellipsoidal clusters. Conventional relational clustering models such as relational fuzzy c-means or relational fuzzy c-medoids produce bad results for this family of relational data, because they do not consider the cluster shape. In this paper, we develop a Gustafson Kessel model where the cluster centers are medoids. For relational data, the scatter matrices and the matrix distances are locally computed using triangulation. The resulting RGKMdd algorithm produces very good results for the family of relational data specified above

The original ant system algorithm is simplified leading to a generalized ant colony optimization algorithm that can be used to solve a wide variety of discrete optimization problems. It is shown how objective function based clustering models such as hard and fuzzy c-means can be optimized using particular extensions of this simplified ant optimization algorithm. Experiments with artificial and real datasets show that ant clustering produces better results than alternating optimization because it is less sensitive to local extrema. © 2005 Wiley Periodicals, Inc. Int J Int Syst 20: 1233–1251, 2005.

Histograms and chi-squared tests for independence are well defined for discrete data. In order to apply these methods to continuous data, some kind of discretization is necessary. A standard way of discretizing data is to use equally spaced (crisp) intervals. In this paper, this crisp discretization is modified to a fuzzy discretization. With this fuzzy discretization, definitions of fuzzy histograms and fuzzy chi-squared tests for independence are achieved. Six experiments indicate that these fuzzy data analysis methods outperform their crisp relatives in terms of smoothness, robustness against outliers, sensitivity for the position of data clusters, and sensitivity for the number of discretization bins.

This paper addresses the problem of optimizing the schedule of logistic processes using genetic algorithms and fuzzy decision making. We consider here the problem of dynamically assign components to orders and choose the solution that is able to deliver more orders at the correct date, respecting at the same time the priority degree of the orders. A compromise between these conflicting goals is achieved by using a genetic algorithm to optimize a fuzzy weighted function. The simulation results show that the proposed genetic algorithm evolved with fuzzy optimization presents good results for this type of problems.

The objective functions for nonlinear projection and clustering are combined and lead to the definition of fuzzy nonlinear projection. Conventional nonlinear projection preserves topologies well, but produces bad results for multiple manifolds. Conventional clustering can discover complex cluster shapes, but the geometry has to be specified in advance. Fuzzy nonlinear projection avoids these drawbacks of projection and clustering methods. It produces both good projections and good partitions for data sets that contain arbitrarily shaped multiple nonlinear manifolds.

This paper addresses the problem of optimizing logistic processes that can be modeled as a birth-and-death process. A fuzzy decision making algorithm is proposed to assign components to orders, which is a common task in a large number of logistic processes. The dynamic assignment of components to orders is the key issue in optimizing logistic processes. This paper proposes several criteria for this optimization. These criteria are combined using weighted fuzzy aggregation in a fuzzy decision making environment. First, a simple but illustrative example shows that the proposed techniques can be applied with good results to this type of processes. Then, the proposed method is applied to a real-world logistic process at Fujitsu-Siemens Computers.

This paper addresses the problem of optimizing logistic processes that can be modeled as birth and death processes. A fuzzy decision making algorithm is proposed to dynamically assign components to orders. This algorithm tries to minimize the overall delivery delays and, at the same type, prefers orders with high priorities. A compromise between these conflicting goals is achieved by using weighted fuzzy aggregation. A logistic example shows the effectiveness of the proposed fuzzy method.

In the field of road traffic management, fuzzy techniques have already been used for traffic control. In this paper, we use fuzzy methods for traffic data analysis. The results of the data analysis are classification and prediction systems. Our work is focused on fuzzy clustering methods. The known clustering models are extended to: constrained prototypes, the use of a mix of different prototypes for one data set, partial supervision of the clustering, and the estimation of the number of clusters by cluster merging. Two successful application examples are given. The first one is the classification of traffic jam on a German autobahn, and the second application is a long-term prediction of traffic volume.

The key issue of a multi agent system is the decentralized control and optimization of a given global system which is normally of large scale. In this connection features like stability, controllability, and observability are important aspects. The paper deals with control structures of decentralized agent driven hybrid systems. In addition some stability features of decentralized hybrid systems are discussed. An example of agent control of two coupled heaters shows that even stable systems may show periodic behavior.

Alternating cluster estimation (ACE) is a generalized clustering model. Relational ACE is a modification of ACE that can be used to cluster data which do not possess a clear numerical representation, but for which a meaningful relation matrix can be defined. For text data sets we define (pairwise) relation matrices based on the Levenshtein string distance (1966). Relational ACE with Levenshtein distances is applied to four different texts. The cluster centers represent typical words in the texts, so this algorithm can be used to automatically determine keywords.

Many clustering models define good clusters as extrema of objective functions. Optimization of these models is often done using an alternating optimization (AO) algorithm driven by necessary conditions for local extrema. We abandon the objective function model in favor of a generalized model called alternating cluster estimation (ACE). ACE uses an alternating iteration architecture, but membership and prototype functions are selected directly by the user. Virtually every clustering model can be realized as an instance of ACE. Out of a large variety of possible instances of non-AO models, we present two examples: 1) an algorithm with a dynamically changing prototype function that extracts representative data and 2) a computationally efficient algorithm with hyperconic membership functions that allows easy extraction of membership functions. We illustrate these non-AO instances on three problems: a) simple clustering of plane data where we show that creating an unmatched ACE algorithm overcomes some problems of fuzzy c-means (FCM-AO) and possibilistic c-means (PCM-AO); b) functional approximation by clustering on a simple artificial data set; and c) functional approximation on a 12 input 1 output real world data set. ACE models work pretty well in all three cases.

Defuzzification is used to transform fuzzy inference results into crisp output. The standard defuzzification methods fail in some applications. It is, therefore, important to select appropriate defuzzification methods depending on the application. This paper presents some of the most important defuzzification methods and investigates their properties. With three application examples, it illustrates how to select appropriate defuzzification methods using application specific properties.