Artificial Intelligence XXXV

The paper introduces the concept of $$\varPhi $$ -data, data that is a proxy for some underlying data that offers advantages of data privacy and security while at the same time allowing particular data mining operations without requiring data owner participation once the proxy has been generated. The nature of the proxy representation is dependent on the nature of the desired data mining to be undertaken. Secure collaborative clustering is considered where the $$\varPhi $$ -data is in the form of a Super Secure Chain Distance Matrices (SSCDM) encrypted using a proposed Multi-User Order Preserving Encryption (MUOPE) scheme. SSCDMs can be produced with respect to horizontal and vertical data partitioning. The DBSCAN clustering algorithm is adopted for illustrative and evaluation purposes. The results indicate that the proposed solution is efficient and produces comparable clustering configurations to those produced using an unencrypted, “standard”, algorithm; while maintaining data privacy and security.

Rule based systems are an important class of computer languages. The brain, and more recently neuromorphic systems, is based on neurons. This paper describes a mechanism that converts a rule based system, specified by a user, to spiking neurons. The system can then be run in simulated neurons, producing the same output. The conversion is done making use of binary cell assemblies, and finite state automata. The binary cell assemblies, eventually implemented in neurons, implement the states. The rules are converted to a dictionary of facts, and simple finite state automata. This is then cached out to neurons. The neurons can be simulated on standard simulators, like NEST, or on neuromorphic hardware. Parallelism is a benefit of neural system, and rule based systems can take advantage of this parallelism. It is hoped that this work will support further exploration of parallel neural and rule based systems, and support further work in cognitive modelling and cognitive architecture.

Siamese Neural Networks (SNNs) are deep metric learners that use paired instance comparisons to learn similarity. The neural feature maps learnt in this way provide useful representations for classification tasks. Learning in SNNs is not reliant on explicit class knowledge; instead they require knowledge about the relationship between pairs. Though often ignored, we have found that appropriate pair selection is crucial to maximising training efficiency, particularly in scenarios where examples are limited. In this paper, we study the role of informed pair selection and propose a 2-phased strategy of exploration and exploitation. Random sampling provides the needed coverage for exploration, while areas of uncertainty modeled by neighbourhood properties of the pairs drive exploitation. We adopt curriculum learning to organise the ordering of pairs at training time using similarity knowledge as a heuristic for pair sorting. The results of our experimental evaluation show that these strategies are key to optimising training.

In recent years, some impressive AI systems have been built that can play games and answer questions about large quantities of data. However, we are still a very long way from AI systems that can think and learn in a human-like way. We have a great deal of information about how the brain works and can simulate networks of hundreds of millions of neurons. So it seems likely that we could use our neuroscientific knowledge to build brain-inspired artificial intelligence that acts like humans on similar timescales. This paper describes an AI system that we have built using a brain-inspired network of artificial spiking neurons. On a word recognition and colour naming task our system behaves like human subjects on a similar timescale. In the longer term, this type of AI technology could lead to more flexible general purpose artificial intelligence and to more natural human-computer interaction.

In organisations with a large mobile workforce there is a need to improve the operational efficiency of the engineers who form the mobile workforce. This improvement can lead to significant savings in operational costs and a corresponding increase in revenue. The operational efficiency of the engineers can be improved by optimising the geographic areas within which the engineers operate. This process is known as Work Area Optimization and it is a subdomain of Workforce Optimization. In this paper, we will present a Hybrid Genetic Algorithm where we will use Deep Neural Networks to generate prior knowledge about the Work Area Optimization problem and use this knowledge to generate improved initial estimates which in turn improves the performance of an existing Genetic Algorithm that does Work Area Optimization. We will also compare our approach with prior knowledge generated with the help of human experts with years of experience in the field. We show that our new approach is as good as or better in generating the prior knowledge when compared to human experts.

Autonomous controllers are highly expertise entities that integrate the sensing-planning-act cycle to operate robotic platforms in unaffordable environments. Its complexity usually makes them to be focused on a single robotic platform which is ostensibly inefficient. The Teleo-Reactive EXecutive (T-REX) is an autonomous controller envisaged as a multi-agent architecture where sensing, planning and execution are interleaved on a single agent. In this paper, we present a T-REX executive module to manage the execution cycle of actions during the planning phase. Our executive module, called GER, aims to state generic execution policies which make a T-REX controller turns into a heterogeneous entity able to operate over any robotic domain. The experimental section demonstrates that GER allows current T-REX architectures, such as GOAC, to manage different robotic domains as Unmanned Aerial Vehicles (UAV) or Unmanned Ground Vehicles (UGV).

Considering Markovian Decision Processes (MDPs), the meaning of an optimal policy depends on the optimality criterion chosen. The most common approach is to define the optimal policy as the one that maximizes the sum of discounted rewards. The intuitive alternative is to maximize the average reward per step. The former has strong convergence guarantees but suffers from the dependency on a discount factor. The latter has the additional inconvenience of being insensitive to different policies with equivalent average. This paper analyzes the impact of such different criteria on a series of experiments, and then provides a threshold for the discount factor in order to ensure average optimality for discounted-optimal policies in the deterministic case.

Mars exploration is an ongoing researching topic mainly due to the technological breakthroughs in robotic platforms. Space agencies as NASA, are considering future Mars explorations where multi-robot teams cooperate to maximize the scientific return. In this regard, we present a cooperative team formed by a Unmanned Aerial Vehicle (UAV) and a Unmanned Ground Vehicle (UGV) to autonomously perform a Mars exploration. We develop a strategical path planner to compute a route plan for the UGV-UAV team to reach all the target points of the exploration. The key problems that we have considered in Mars explorations for the UGV-UAV team are: the UAV energy constraints and the UGV functionality constraints. Our strategical path planner models the UGV as a moving charging station which will carry the UAV through secure locations close to the target points locations, and the UAV will visit the target points using the UGV as a recharging station. Our solution has been tested in several scenarios and the results demonstrate that our approach is able to carry out a coordinated plan in a local optimal mission time on a real Mars terrain.

The paper proposes a new ensemble classifier using Gene Expression Programming as the induction engine. The approach aims at predicting unknown class labels for datasets with concept drift. For constructing the proposed ensemble we use the two-level scheme where at the lower level base classifiers are induced and at the upper level, the meta-classifier is produced. The classification process is controlled by the well-known early drift detection mechanism. To validate the approach computational experiment has been carried out. Its results confirmed that the proposed classifier performs well.

Dynamic business processes are challenged by constant changes due to unstable environments, unexpected incidents and difficult to predict behaviours. In industry areas like customer support, complex incidents can be regarded as instances of a dynamic process since there can be no static planning against their unique nature. Support engineers will work with any means at their disposal to solve any emerging case and define a custom prioritization strategy, to achieve the best possible result. To assist with this, in this paper we describe a novel workflow application to address the tasks of high solution accuracy and shorter prediction resolution time. We describe how workflows can be generated to assist experts and how our solution can scale over time to produce domain-specific reusable cases for similar problems. Our work is evaluated using data from 5000 workflows from the automotive industry.

Reinforcement learning has shown great potential in generalizing over raw sensory data using only a single neural network for value optimization. There are several challenges in the current state-of-the-art reinforcement learning algorithms that prevent them from converging towards the global optima. It is likely that the solution to these problems lies in short- and long-term planning, exploration and memory management for reinforcement learning algorithms. Games are often used to benchmark reinforcement learning algorithms as they provide a flexible, reproducible, and easy to control environment. Regardless, few games feature a state-space where results in exploration, memory, and planning are easily perceived. This paper presents The Dreaming Variational Autoencoder (DVAE), a neural network based generative modeling architecture for exploration in environments with sparse feedback. We further present Deep Maze, a novel and flexible maze engine that challenges DVAE in partial and fully-observable state-spaces, long-horizon tasks, and deterministic and stochastic problems. We show initial findings and encourage further work in reinforcement learning driven by generative exploration.

Virtualisation has revolutionised computing, enabling applications to be quickly provisioned and deployed compared to traditional systems and ensuring that client applications have an ongoing quality of service, with dynamic resourcing in response to demand. However, this requires the use of performance metrics, to recognise current or evolving resourcing situations and ensure timely reprovisioning or redeployment. Associated monitoring systems should thus be aware of not only individual metric behaviours but also of the relationship between related metrics so that system alarms can be triggered when the metrics fall outside normal operational parameters. We here consider multivariate approaches, namely analysis of correlation structure and multivariate exponentially weighted moving averages (MEWMA), for detecting abnormalities in cloud performance data with a view to timely intervention.

The most popular pathfinding approach used in video games is the A* algorithm. This paper looks at depth-first search to determine if, with modifications, it can be made into a viable alternative. Proposed is a method of directing a search, which utilises a scoring system. Tests conducted on randomly generated maps of varying sizes showed that Directed Recursion Search calculated near-optimal paths in less time, by expanding fewer nodes, and with less memory required to store the paths, than A* or depth-first search.

Individual and Societal Trust in AI will determine the scope of application and level of adoption of AI technology. A model of trust is proposed with its partial implementation as a numerical simulation. It is shown by simulation that the introduction of multi-agent dynamics of the sort observed in the case studies we describe has a significant impact on the behaviour of the system. The contributions of this paper are; a new model of trust in AI systems, the partial realisation of the model as a simulator and the results of the preliminary experiments over our simulation.

In this paper, we explain how to benefit from the reasoning capabilities of existential rules for modelling an MCDM problem as an inconsistent knowledge base. The repairs of this knowledge base represent the maximally consistent point of views and inference strategies can be used for decision making.

This paper explores whether a simple grammar-based metric can accurately predict human opinion of machine-generated song lyrics squality. The proposed metric considers the percentage of words written in natural English and the number of grammatical errors to rate the quality of machine-generated lyrics. We use a state-of-the-art Recurrent Neural Network (RNN) model and adapt it to lyric generation by re-training on the lyrics of 5,000 songs. For our initial user trial, we use a small sample of songs generated by the RNN to calibrate the metric. Songs selected on the basis of this metric are further evaluated using “Turing-like” tests to establish whether there is a correlation between metric score and human judgment. Our results show that there is strong correlation with human opinion, especially at lower levels of song quality. They also show that 75% of the RNN-generated lyrics passed for human-generated over 30% of the time.

Decision tree is a simple but powerful learning technique that is considered as one of the famous learning algorithms that have been successfully used in practice for various classification tasks. They have the advantage of producing a comprehensible classification model with satisfactory accuracy levels in several application domains. In recent years, the volume of data available for learning is dramatically increasing. As a result, many application domains are faced with a large amount of data thereby posing a major bottleneck on the computability of learning techniques. There are different implementations of the decision tree using different techniques. In this paper, we theoretically and experimentally study and compare the computational power of the most common classical top-down decision tree algorithms (C4.5 and CART). This work can serve as part of review work to analyse the computational complexity of the existing decision tree classifier algorithm to gain understanding of the operational steps with the aim of optimizing the learning algorithm for large datasets.

The need for qualified people is growing exponentially, requiring limited resources allocated to education/training to be used most efficiently. However some problems can occur: (1) relying on learning theories, it is crucial to improve the learning process and mitigate the issues that may arise from technologically enhanced learning environments; (2) each student presents a particular way of assimilating knowledge, i.e. his/her learning procedure. It’s essential that these systems adapt to the learning preferences of the students. In the present study, we propose an intelligent learning system able to monitor the patterns of students’ behaviour during e-assessments, to support the teaching procedure within school environments. Results show that there are still mechanisms that can be explored to understand better the complex relationship between human behaviour, attention, and assessment which could be used for the implementation of better learning strategies. These results may be crucial for improving learning systems in an e-learning environment and for predicting students’ behaviour in an exam, based on their interaction with technological devices.

This work addresses the recognition of human motion exercises using 3D-skeleton-data and Neural Networks (NN). The examined dataset contains 16 gymnastic motion exercises (e.g. squats, lunges) executed from 21 subjects and captured with the second version of the MicrosoftTM Kinect sensor (Kinect v2). The NN was trained with eight datasets from eight subjects and tested with 13 unknown datasets. The investigation in this work focuses on the configuration of NNs for human motion recognition. The authors will conclude that a backpropagation NN consisting of 100 neurons, three hidden layers, and a learning rate of 0.001 reaches the best accuracy with 93.8% correct.

Dynamic planning is a hot topic in autonomous computing. This work presents a novel approach of simulating swarm computing behaviour in a sandbox environment where swarms of robots are challenged to fight against each other with a goal of “conquering” any environment bases. Swarm strategies are being used which are decided, modified and applied at run time. Autonomous swarm agents seem surprisingly applicable to several problems where combined artificial intelligence agents are challenged to generate innovative solutions and evaluate them prior to proposing or adopting the best possible one. This work is applicable in areas where AI agents should make selections close to real time within a range of available options under a multi-constraint, multi-objective mission environment. Relevance to Business Process workflows is also presented and documented.

Over the past decades, football (soccer) has continued to draw more and more attention from people all over the world. Meanwhile, the appearance of the internet led to a rapidly growing market for online bookmakers, companies which offer sport bets for specific odds. With numerous matches every week in dozens of countries, football league matches hold enormous potential for developing betting strategies. In this context, a betting strategy beats the bookmaker if it generates positive average profits over time. In this paper, we developed a data-driven framework for predicting the outcome of football league matches and generating meaningful profits by betting accordingly. Conducting a simulation study based on the matches of the five top European football leagues from season 2013/14 to 2017/18 showed that economically and statistically significant returns can be achieved by exploiting large data sets with modern machine learning algorithms. Furthermore, it turned out that these results cannot be reached with a linear regression model or simple betting strategies, such as always betting on the home team.

The analysis of complex business processes is a challenging topic. Machine learning provides many tools to help with the analysis/understanding of complex data. In this paper we present the application of two types of technique to this domain. First, rule mining techniques to discover relationships between process behaviour and outcomes. Second, a technique presented is one suitable for clustering arbitrary length directed acyclic graphs such as those that represent business process executions. Both cases are presented in the context of a real business process.

Control systems are in general based on the same structure, building blocks and physics-based models of the dynamic system regardless of application, and can be mathematically analyzed w.r.t. stability, robustness and so on given certain assumptions. Machine learning methods (ML), on the other hand, are highly flexible and adaptable methods but are not subject to physic-based models and therefore lack mathematical analysis. This paper presents state of the art results using ML in the control system. Furthermore, a case study is presented where a neural network is trained to mimic a feedback linearizing speed controller for an autonomous ship. The neural network outperforms the traditional controller in case of modeling errors and measurement noise.

Prediction is about making claims on future events based on past information and the current state. Predicting workforce demand for the future can help service organizations adjust their resources and reach their goals of cost saving and enhanced efficiency. In this paper, a use case for a telecom service organization is presented and a framework for predicting workforce demand using neural networks is provided. The experiments were performed with real-world data, and the results were compared against other popular techniques such as linear regression and also moving average which served as a simulation of the technique historically applied manually in the organization. The results show that the accuracy of prediction is improved with the use of neural networks. The technique is being built into a tool that is being tested by the partner telecom organization.

Staff scheduling and planning in a cost effective manner has been a topic of scientific discussion for many years, driven by the need of many organisations to fully and effectively utilise their workforce to meet costumer demand and deliver service. Due to the varying nature of industry sectors, problems often require tailoring for particular business needs and types of work. This paper presents an overview of how a version of this problem was solved in a business with a large field workforce. The automation of this process has proved vital in ensuring that there is the right amount of resources rostered in on each day of the week, transforming a lengthy, manual procedure into an operation of a matter of seconds. The paper discusses how a Simulated Annealing approach was implemented, and provides a comparison of its performance versus a standard Hill Climber. We also include a detailed description of how rules and constraints were incorporated into the work, and what effect these had on rostered attendance.

This paper presents a model for the risk minimisation objective in the Stochastic Vehicle Routing Problem (SVRP). In the studied variant of SVRP, service times and travel times are subject to stochastic events, and a time window is constraining the start time for service task. Required skill levels and task priorities increase the complexity of this problem. Most previous research uses a chance-constrained approach to the problem and their objectives are related to traditional routing costs whilst a different approach was taken in this paper. The risk of missing a task is defined as the probability that the technician assigned to the task arrives at the customer site later than the time window. The problem studied in this paper is to generate a schedule that minimises the maximum of risks and sum of risks over all the tasks considering the effect of skill levels and task priorities. The stochastic duration of each task is supposed to follow a known normal distribution. However, the distribution of the start time of the service at a customer site will not be normally distributed due to time window constraints. A method is proposed and tested to approximate the start time distribution as normal. Moreover, a linear model can be obtained assuming identical variance of task durations. Additionally Simulated Annealing method was applied to solve the problem. Results of this work have been applied to an industrial case of SVRP where field engineering individuals drive to customer sites to provide time-constrained services. This original approach gives a robust schedule and allows organisations to pay more attention to increasing customer satisfaction and become more competitive in the market.

Within any sufficiently expertise-reliant and work-driven domain there is a requirement to understand the similarities between specific work tasks. Though mechanisms to develop similarity models for these areas do exist, in practice they have been criticised within various domains by experts who feel that the output is not indicative of their viewpoint. In field service provision for telecommunication organisations, it can be particularly challenging to understand task similarity from the perspective of an expert engineer. With that in mind, this paper demonstrates a similarity model developed from text recorded by engineer’s themselves to develop a metric directly indicative of expert opinion. We evaluate several methods of learning text representations on a classification task developed from engineers’ notes. Furthermore, we introduce a means to make use of the complex and multi-faceted aspect of the notes to recommend additional information to support engineers in the field.

In recent decades, multimedia data have been commonly generated and used in various domains, such as in healthcare and social media due to their ability of capturing rich information. But as they are unstructured and separated, how to fuse and integrate multimedia datasets and then learn from them effectively have been a main challenge to machine learning. We present a novel generalised decision level ensemble method (GDLEM) that combines the multimedia datasets at decision level. After extracting features from each of multimedia datasets separately, the method trains models independently on each media dataset and then employs a generalised selection function to choose the appropriate models to construct a heterogeneous ensemble. The selection function is defined as a weighted combination of two criteria: the accuracy of individual models and the diversity among the models. The framework is tested on multimedia data and compared with other heterogeneous ensembles. The results show that the GDLEM is more flexible and effective.

Earnings manipulation and accounting fraud leads to reduced firm valuation in the long run and a public distrust in the company and its management. Yet, manipulation of accruals to hide liabilities and inflate earnings has been a long-standing fraudulent conduct amongst many listed firms. As auditing is time consuming and restricted to a sample of entries, fraud is either not detected or detected belatedly. We believe that supervised machine learning models can be used to determine high risk firms early enough for auditing by the regulator. We also discuss the anomaly detection unsupervised learning methodology. Since the proportion of manipulators is much lower than the non-manipulators, the biggest challenge in predicting earnings manipulation is the imbalance in the data leading to biased results for conventional statistical models. In this paper, we build ensemble models to detect accrual manipulation by borrowing theory from the seminal work done by Beneish. We also showcase a novel simulation-based sampling technique to efficiently handle imbalanced dataset and illustrate our results on data from listed Indian firms. We compare existing ensemble models establishing the superiority of fairly simple boosting models whilst commenting on the shortfall of area under ROC curve as a performance metric for imbalanced datasets. The paper makes two major contributions: (i) a functional contribution of suggesting an easily deployable strategy to identify high risk companies; (ii) a methodological contribution of suggesting a simulation-based sampling approach that can be applied in other cases of highly imbalanced data for utilizing the entire dataset in modeling.

Aspect-level sentiment analysis of customer feedback data when done accurately can be leveraged to understand strong and weak performance points of businesses and services and also formulate critical action steps to improve their performance. In this work we focus on aspect-level sentiment classification studying the role of opinion context extraction for a given aspect and the extent to which traditional and neural sentiment classifiers benefit when trained using the opinion context text. We introduce a novel method that combines lexical, syntactical and sentiment knowledge effectively to extract opinion context for aspects. Thereafter we validate the quality of the opinion contexts extracted with human judgments using the BLEU score. Further we evaluate the usefulness of the opinion contexts for aspect-sentiment analysis. Our experiments on benchmark data sets from SemEval and a real-world dataset from the insurance domain suggests that extracting the right opinion context combining syntactical with sentiment co-occurrence knowledge leads to the best aspect-sentiment classification performance. From a commercial point of view, accurate aspect extraction, provides an elegant means to identify “pain-points” in a business. Integrating our work into a commercial CX platform ( https://www.sentisum.com/ ) is enabling the company’s clients to better understand their customer opinions.

In this paper, we present a non-deterministic strategy for searching for optimal number of trees (NoTs) hyperparameter in Random Forest (RF). Hyperparameter tuning in Machine Learning (ML) algorithms optimizes predictability of an ML algorithm and/or improves computer resources utilization. However, hyperparameter tuning is a complex optimization task and time consuming. We set up experiments with the goal of maximizing predictability, minimizing NoTs and minimizing time of execution (ToE). Compared to the deterministic algorithm, e-greedy and default configured RF, this research’s non-deterministic algorithm recorded an average percentage accuracy (acc) of approximately 98%, NoTs percentage average improvement of 29.39%, average ToE improvement ratio of 415.92 and an average improvement of 95% iterations. Moreover, evaluations using Jackknife Estimation showed stable and reliable results from several experiment runs of the non-deterministic strategy. The non-deterministic approach in selecting hyperparameter showed a significant acc and better computer resources (i.e. cpu and memory time) utilization. This approach can be adopted widely in hyperparameter tuning, and in conserving utilization of computer resources i.e. green computing.

This paper describes the use of a genetic algorithm to design a website, according to principles of clarity, symmetry, golden ratio and image size. The website’s logo is used to calculate a matching colour scheme. Results indicate that local maxima can be a problem but that with the right weighting of the fitness function, a pleasing design can be achieved.Such a program could be used when designing large numbers of websites; when a website has to be re-designed regularly to match changing content; or to provide a starting point for human website designers or users of interactive genetic algorithms to improve.

This paper describes applications of BOBBIN, a multi-agent system based on a blackboard architecture, to supporting smart cities through regulated information sharing and pattern recognition. The first application uses the knowledge-based processing provided by a blackboard system to ensure that personal information is made available to all and only those who are entitled to see it, thus overcoming the objections raised by the UK Supreme Court to a recently proposed information sharing scheme. The second application extends the first to enable pattern recognition over appropriately regulated information about individuals, with the goal of identifying criminal offences and other patterns of behavior. The final application extends the second to deal with heterogeneous agents, public and private. Each application is illustrated with a scenario.These three approaches can provide benefits in supporting the growth of intelligent public services and appropriate information sharing within a smart city.

Lego Mindstorms has delivered low-cost amateur robotics to the public, where anyone can easily modify and develop new systems and extensions to extend its capabilities. However, no one has previously attempted to link a 4th generation agent-oriented language such as NetLogo with Mindstorms robots in order to provide an agent-oriented development environment along with simulation and modelling capabilities. This paper describes the development of middleware which can be used to control a Mindstorms robot via a NetLogo model which provides body-syntonic capabilities for real-time sensor feeds and robot commands to make and enact decisions. A couple of example NetLogo models to demonstrate the capabilities of this system (line-following ability and subsumption architecture roaming) have been developed and are described in this paper.

In this paper, we introduce a dynamic variant of the Location-Allocation problem: Dynamic Location-Allocation Problem (DULAP). DULAP involves the location of facilities to service a set of customer demands over a defined horizon. To evaluate a solution to DULAP, we propose two holistic metric approaches: Static and Dynamic Approach. In the static approach, a solution is evaluated with the assumption that customer locations and demand remain constant over a defined horizon. In the dynamic approach, the assumption is made that customer demand, and demographic pattern may change over the defined horizon. We introduce a stochastic model to simulate customer population and distribution over time. We use a Genetic Algorithm and Population-Based Incremental Learning algorithm used in previous work to find robust and satisfactory solutions to DULAP. Results show the dynamic approach of evaluating a solution finds good and robust solutions.

Customer adoption of innovation is a multi-disciplinary research area which has been extensively researched. However, identifying the most important variables that affect the adoption process remains unattended. In this paper, we propose Fuzzy TOPSIS method to rank variables that affect the Innovator Group customers. With an illustrative example we explain the method’s applicability and we conclude by discussing implications for marketing research and healthy food industry.

Rail transportation is an important part of the transport infrastructure that supports modern advanced economies. Both public and private companies are highly concerned on how travel patterns, vehicle-passenger behaviours and other relevant phenomena such as weather affect their performance. Usually any travel network can be remarkably expensive to build and swiftly gets saturated after its construction and any subsequent upgrades. We propose suitable workflow monitoring methods for developing efficient performance measures for the rail industry using business process workflow pattern analysis based on Case-based Reasoning (CBR) combined with standard Data Mining methods. The approach focuses on both data preparation and cleaning and integration of data applied to a real industrial case study. Preliminary results of this work are promising against the complexity of the data and can scale on demand while showing they can predict to an efficient accuracy. Several modelling experiments are presented, that show that the proposed approach can provide a sound basis for effective and useful analysis of operational sensor data from train Journeys.