Search

In this work, we provide a more consistent alternative for performing value of information (VOI) analyses to address sequential decision problems in reservoir management and generate insights on the process of reservoir decision-making. These sequential decision problems are often solved and modeled as stochastic dynamic programs, but once the state space becomes large and complex, traditional techniques, such as policy iteration and backward induction, quickly become computationally demanding and intractable. To resolve these issues and utilize fewer computational resources, we instead make use of a viable alternative called approximate dynamic programming (ADP), which is a powerful solution technique that can handle complex, large-scale problems and discover a near-optimal solution for intractable sequential decision making. We compare and test the performance of several machine learning techniques that lie within the domain of ADP to determine the optimal time for beginning a polymer flooding process within a reservoir development plan. The approximate dynamic approach utilized here takes into account both the effect of the information obtained before a decision is made and the effect of the information that might be obtained to support future decisions while significantly improving both the timing and the value of the decision, thereby leading to a significant increase in economic performance.

PhD thesis in Information technology Renewable resources are the main energy sources in a smart grid project. In order to ensure the smooth functioning of the smart grid, Information and Communication Technologies (ICT) need to be utilised efficiently. The objective of the SmartNEM project is to effectively utilise the technologies such as Machine Learning, Blockchain and Data Hubs for the aforementioned purpose and at the same time ensure a secured and privacy preserved solution. The data involved in smart grids require high security and it can be sensitive due to the household data which contains personal information. The individuals can be reluctant to share these data due to mistrust and to avoid unnecessary manipulation of the data they provide. In order to overcome this it is necessary to build a trust based framework in which one could ensure data security and data privacy for the data owners to open up their data for data analysis. To achieves this we have proposed an architecture called TOTEM, Token for Controlled Computation, which integrates Blockchain and Big Data technologies. The conventional method of data analysis demands data be moved across the network to the location where the execution happens, however in the TOTEM architecture computational code will be moved to the data owner’s environment where the data is located. The TOTEM is a three layer architecture (Blockchain consortium layer, Storage layer and Computational layer) with two main actors, data provider and data consumer. Data provider provides metadata of the data they own and provide resources for the execution of data. Data consumers will get an opportunity to execute their own code on the data provider´s data. For a controlled computation and to avoid malicious functions an entity called totem is introduced in the architecture. The authorised users should meet the requirements of Totem value for executing their code on the requested data. For live monitoring of the totem value throughout the run time is achieved with the components such as totem manager and updaters in the computational layer. The code must follow a specific format and will undergo preliminary checks with the TOTEM defined SDK and smart contracts deployed by the data providers in the blockchain network. The Extended TOTEM architecture is also proposed to address the additional features when it is needed to combine the results from multiple data providers without sharing the data. This research work focused on the design of the TOTEM architecture and implementation as a proof of concept for the newly introduced components in the architecture. We have also introduced artificial intelligence in the framework to improve core features’ functionality. In the present research, the TOTEM architecture is proposed for the SmartNEM project to utilize the energy data for decision making and figure out the trends or patterns, while maintaining data privacy, data ownership, accountability and traceability. Moreover, the architecture can be extended to other domains such as health, education, etc, where data security and privacy is the key concern in sharing the data.

Extraterrestrial rovers with a general-purpose robotic arm have many potential applications in lunar and planetary exploration. Introducing autonomy into such systems is desirable for increasing the time that rovers can spend gathering scientific data and collecting samples. This work investigates the applicability of deep reinforcement learning for vision-based robotic grasping of objects on the Moon. A novel simulation environment with procedurally-generated datasets is created to train agents under challenging conditions in unstructured scenes with uneven terrain and harsh illumination. A model-free off-policy actor-critic algorithm is then employed for end-to-end learning of a policy that directly maps compact octree observations to continuous actions in Cartesian space. Experimental evaluation indicates that 3D data representations enable more effective learning of manipulation skills when compared to traditionally used image-based observations. Domain randomization improves the generalization of learned policies to novel scenes with previously unseen objects and different illumination conditions. To this end, we demonstrate zero-shot sim-to-real transfer by evaluating trained agents on a real robot in a Moon-analogue facility., Comment: Accepted for publication at the 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) | The source code is available at https://github.com/AndrejOrsula/drl_grasping | The supplementary video is available at https://youtube.com/watch?v=FZSoOkK6VFc

Due to the current COVID-19 virus, which has already been declared a pandemic by the World Health Organization (WHO), we are witnessing the greatest pandemic of the decade. Millions of people are being infected, resulting in thousands of deaths every day across the globe. Even it was difficult for the best healthcare-providing countries could not handle the pandemic because of the strain of treating thousands of patients at a time. The count of infections and deaths is increasing at an alarming rate because of the spread of the virus. We believe that innovative technologies could help reduce pandemics to a certain extent until we find a definite solution from the medical field to handle and treat such pandemic situations. Technology innovation has the potential to introduce new technologies that could support people and society during these difficult times. Therefore, this paper proposes the idea of using drones as a companion to tackle current and future pandemics. Our COROID drone is based on the principle of crowdsourcing sensors data of the public's smart devices, which can correlate the reading of the infrared cameras equipped on the COROID drones. To the best of our knowledge, this concept has yet to be investigated either as a concept or as a product. Therefore, we believe that the COROID drone is innovative and has a huge potential to tackle COVID-19 and future pandemics., Accepted

Traditional decline curve analyses (DCAs), both deterministic and probabilistic, use specific models to fit production data for production forecasting. Various decline curve models have been applied for unconventional wells, including the Arps model, stretched exponential model, Duong model, and combined capacitance-resistance model. However, it is not straightforward to determine which model should be used, as multiple models may fit a dataset equally well but provide different forecasts, and hastily selecting a model for probabilistic DCA can underestimate the uncertainty in a production forecast. Data science, machine learning, and artificial intelligence are revolutionizing the oil and gas industry by utilizing computing power more effectively and efficiently. We propose a data-driven approach in this paper to performing short term predictions for unconventional oil production. Two states of the art level models have tested: DeepAR and used Prophet time series analysis on petroleum production data. Compared with the traditional approach using decline curve models, the machine learning approach can be regarded as” model-free” (non-parametric) because the pre-determination of decline curve models is not required. The main goal of this work is to develop and apply neural networks and time series techniques to oil well data without having substantial knowledge regarding the extraction process or physical relationship between the geological and dynamic parameters. For evaluation and verification purpose, The proposed method is applied to a selected well of Midland fields from the USA. By comparing our results, we can infer that both DeepAR and Prophet analysis are useful for gaining a better understanding of the behavior of oil wells, and can mitigate over/underestimates resulting from using a single decline curve model for forecasting. In addition, the proposed approach performs well in spreading model uncertainty to uncertainty in production forecasting; that is, we end up with a forecast which outperforms the standard DCA methods.

With the advent of ubiquitous sensors and Internet of Things (IoT) applications, research and development initiatives on smart cities are ramping up worldwide. It enables remote monitoring, management, and control of devices and the generation of fresh and actionable insight from huge quantities of real-time data. Real-time noise and emissions monitoring of vehicles remain indispensable in a smart city context. Effective management and control of noise and emissions of vehicles on the road are necessary and possible through analyzing lots of sensor data in real-time to take an actionable insight. To contribute to this, as part of an ongoing effort of the European Union project called ''NEMO: Noise and Emissions Monitoring and Radical Mitigation'', in this paper, we present the design and development of an IoT-based real-time noise and emissions monitoring system for vehicles in a smart city context. Real-world sensor data of the vehicles in some European cities are collected during the pilot tests. We have developed a complete application for infrastructure managers and analysts to monitor the sensor data related to noise and emissions of vehicles in real-time. The data of the individual road vehicles and trains in selected EU cities and from trains on a track in the Netherlands are collected in the cloud and analyzed with artificial intelligence (AI) algorithms for classification such as high emitter, medium emitter, and normal emitters. We present the development of a complete software solution that can be integrated with existing intelligent transportation systems in smart cities. Finally, we report the initial vehicle classification results from the Rotterdam (Netherlands) pilot test as a representative example for the NEMO monitoring system.

Petrophysical interpretation and optimal correlation extraction of different measurements require accurate well log depth matching. We have developed a supervised multimodal machine learning alternative for the task of simultaneously matching raw logging while drilling and electrical wireline logging logs. Seven one-dimensional convolutional neural networks are trained using different log measurements: gamma-ray, resistivity, P- and S-wave sonic, density, neutron and photoelectric factors, and their depth shift estimates are aggregated using different multimodal late fusion strategies. We test the late fusion average, late fusion weighted average, late fusion with linear and nonlinear learners and model-level fusion. Depth matching results using the different fusion strategies applied to two unseen wells are compared using visual inspection and the mean Pearson correlation. All models perform well, increasing the correlation after depth matching. Late fusion weighted average achieves the highest scores for all log types. The late fusion weighted average results are compared to a cross-correlation user-assisted workflow and manual depth matching for validation. In general, the convolutional neural network fused method exhibits a lower performance than the traditional methods. For one of the wells, the cross-correlation shows higher correlation values than the other methods but for the second well the manual depth match performs best. However, the differences in Pearson correlation values are small ranging from 0.01 to 0.1. The manual depth match performs very well for the sonic logs, which tend to require slightly larger depth shifts than the other measurements, thus a common depth shift might not always be suitable. Although our convolutional neural network fused approach is limited to estimating bulk shifts and uses constant fusion weights, its performance is similar to that of more time-consuming methods. Our approach might be substantially improved by including dynamic shifts (stretch/squeeze) and depth-dependent fusion weights via long-short-term memory recurrent neural networks. Automated well log depth matching: Late fusion multimodal deep learning

Most visual odometry (VO) and visual simultaneous localization and mapping (VSLAM) systems rely heavily on robust keypoint detection and matching. With regards to images taken in the underwater environment, phenomena like shallow water caustics and/or dynamic objects like fishes can lead to the detection and matching of unreliable (unsuitable) keypoints within the visual motion estimation pipeline. We propose a plug-and-play keypoint rejection system that rejects keypoints unsuitable for tracking in order to obtain a robust visual ego-motion estimation. A convolutional neural network is trained in a supervised manner, with image patches having a detected keypoint in its center as input and the probability of such a keypoint suitable for tracking and mapping as output. We provide experimental evidence that the system prevents to track unsuitable keypoints in a state-of-the-art VSLAM system. In addition we evaluated several strategies aimed at increasing the inference speed of the network for real-time operations.

skrl is an open-source modular library for reinforcement learning written in Python and designed with a focus on readability, simplicity, and transparency of algorithm implementations. Apart from supporting environments that use the traditional OpenAI Gym interface, it allows loading, configuring, and operating NVIDIA Isaac Gym environments, enabling the parallel training of several agents with adjustable scopes, which may or may not share resources, in the same execution. The library's documentation can be found at this https URL and its source code is available on GitHub at url{this https URL.

Summary The Assisted Cement Log Interpretation Project has used machine learning (ML) to create a tool that interprets cement logs by predicting a predefined set of annular condition codes used in the cement log interpretation process. The development of a cement log interpretation tool speeds up the log interpretation process and enables expert knowledge to be efficiently shared when training new professionals. By using high-quality and consistent training data sets, the project has trained a model that will support unbiased and consistent interpretations over time. The tool consists of a training and a prediction tool integrated with cased-hole logging interpretation software. By containerizing the code using an “API First” design principle (API: application programming interface), the applicability of this add-on tool is broad. The ML model is trained using selected and engineered features from cement logs, and the tool predicts an annular condition code according to the cement classification system for each depth segment in the log. The interpreters can easily fetch a complete cement log interpretation prediction for the log and use that as a template for their final interpretation. The ML model can easily be retrained with new data sets to improve accuracy even further. To improve cement log interpretation consistency in the industry, the code will be made available as open source.

Every time we interact with online digital services, we generate large amounts of data that reveal our shopping habits, social interactions, and much more. We refer to these data collectively as the user-generated digital footprint (UGDF). Today, there is growing interest in using UGDF data as an alternative to conventional financial data in building consumer credit models—UGDF models. Unfortunately, we also observe a hype where the models’ business deployability is reduced to simplistic technical metrics, namely, the model’s prediction accuracy. This study argues that this is a misleading oversimplification of the financial sector’s business realities as it ignores vital dimensions such as the model’s economic viability. Therefore, we develop a framework for evaluating the business deployability of UGDF models for consumer credit using a design science research methodology. The framework is composed of seven criteria: Data accessibility, data coverage, data timeliness, data authenticity, cost of deployment, interpretability, and compliance. A framework for evaluating the business deployability of digital footprint based models for consumer credit

Importance Early identification of cerebral palsy (CP) is important for early intervention, yet expert-based assessments do not permit widespread use, and conventional machine learning alternatives lack validity. Objective To develop and assess the external validity of a novel deep learning–based method to predict CP based on videos of infants’ spontaneous movements at 9 to 18 weeks’ corrected age. Design, Setting, and Participants This prognostic study of a deep learning–based method to predict CP at a corrected age of 12 to 89 months involved 557 infants with a high risk of perinatal brain injury who were enrolled in previous studies conducted at 13 hospitals in Belgium, India, Norway, and the US between September 10, 2001, and October 25, 2018. Analysis was performed between February 11, 2020, and September 23, 2021. Included infants had available video recorded during the fidgety movement period from 9 to 18 weeks’ corrected age, available classifications of fidgety movements ascertained by the general movement assessment (GMA) tool, and available data on CP status at 12 months’ corrected age or older. A total of 418 infants (75.0%) were randomly assigned to the model development (training and internal validation) sample, and 139 (25.0%) were randomly assigned to the external validation sample (1 test set). Exposure Video recording of spontaneous movements. Main Outcomes and Measures The primary outcome was prediction of CP. Deep learning–based prediction of CP was performed automatically from a single video. Secondary outcomes included prediction of associated functional level and CP subtype. Sensitivity, specificity, positive and negative predictive values, and accuracy were assessed. Results Among 557 infants (310 [55.7%] male), the median (IQR) corrected age was 12 (11-13) weeks at assessment, and 84 infants (15.1%) were diagnosed with CP at a mean (SD) age of 3.4 (1.7) years. Data on race and ethnicity were not reported because previous studies (from which the infant samples were derived) used different study protocols with inconsistent collection of these data. On external validation, the deep learning–based CP prediction method had sensitivity of 71.4% (95% CI, 47.8%-88.7%), specificity of 94.1% (95% CI, 88.2%-97.6%), positive predictive value of 68.2% (95% CI, 45.1%-86.1%), and negative predictive value of 94.9% (95% CI, 89.2%-98.1%). In comparison, the GMA tool had sensitivity of 70.0% (95% CI, 45.7%-88.1%), specificity of 88.7% (95% CI, 81.5%-93.8%), positive predictive value of 51.9% (95% CI, 32.0%-71.3%), and negative predictive value of 94.4% (95% CI, 88.3%-97.9%). The deep learning method achieved higher accuracy than the conventional machine learning method (90.6% [95% CI, 84.5%-94.9%] vs 72.7% [95% CI, 64.5%-79.9%]; P Conclusions and Relevance In this prognostic study, a deep learning–based method for predicting CP at 9 to 18 weeks’ corrected age had predictive accuracy on external validation, which suggests possible avenues for using deep learning–based software to provide objective early detection of CP in clinical settings.

Understanding how niche-based and neutral processes contribute to the spatial varia-tion in plant–pollinator interactions is central to designing effective pollination con-servation schemes. Such schemes are needed to reverse declines of wild bees and other pollinating insects, and to promote pollination services to wild and cultivated plants. We used data on wild bee interactions with plants belonging to the four tribes Loteae, Trifolieae, Anthemideae and either spring- or summer-flowering Cichorieae, sampled systematically along a 682 km latitudinal gradient to build models that allowed us to 1) predict occurrences of pairwise bee–flower interactions across 115 sampling locations, and 2) estimate the contribution of variables hypothesized to be related to niche-based assembly structuring processes (viz. annual mean temperature, landscape diversity, bee sociality, bee phenology and flower preferences of bees) and neutral processes (viz. regional commonness and dispersal distance to conspecifics). While neutral processes were important predictors of plant–pollinator distributions, niche-based processes were reflected in the contrasting distributions of solitary bee and bumble bees along the temperature gradient, and in the influence of bee flower preferences on the distri-bution of bee species across plant types. In particular, bee flower preferences separated bees into three main groups, albeit with some overlap: visitors to spring-flowering Cichorieae; visitors to Anthemideae and summer-flowering Cichorieae; and visitors to Trifolieae and Loteae. Our findings suggest that both neutral and niche-based pro-cesses are significant contributors to the spatial distribution of plant–pollinator inter-actions so that conservation actions in our region should be directed towards areas: Page 2 of 11near high concentrations of known occurrences of regionally rare bees; in mild climatic conditions; and that are surrounded by heterogenous landscapes. Given the observed niche-based differences, the proportion of functionally distinct plants in flower-mixes could be chosen to target bee species, or guilds, of conservation concern. Keywords: ecological networks, machine learning, plant–pollinator interactions, spatial, wild bees Neutral processes related to regional bee commonness and dispersal distances are important predictors of plant–pollinator networks along gradients of climate and landscape conditions

Ubiquitous sensors and Internet of Things (IoT)technologies have revolutionized the sports industry, providing new methodologies for planning, effective coordination of training, and match analysis post-game. New methods, including machine learning, image, and video processing, have been developed for performance evaluation, allowing the analyst to track the performance of a player in real-time. Following FIFA’s 2015 approval of electronics performance and tracking system during games, performance data of a single player or the entire team is allowed to be collected using GPS-based wearables. Data from practice sessions outside the sporting arena is being collected in greater numbers than ever before. Realizing the significance of data in professional soccer, this paper presents video analytics, examines recent state-of-the-art literature in elite soccer, and summarizes existing real-time video analytics algorithms. We also discuss real-time crowdsourcing of the obtained data, tactical and technical performance, distributed computing, and its importance in video analytics and propose a future research perspective.

Molecular simulations of biological membranes and proxies thereof are entering a new era characterized by several key aspects. Progress starts with the realization that the outcome of the simulations can only be as good as the underlying force field, and we actually need to know precisely how good or bad the results are. Therefore, standardized procedures for data quality evaluation are being established and will be applied to biomembrane simulations available in the literature. This provides the necessary basis and impetus for new force field development. Here, we propose the systematic buildup of physically well-justified models that effectively account for the electronic polarization effects for all components of the biomembrane systems in aqueous environments. Such a massive task can only be achieved within a reasonable time scale by applying automated parametrization tools. acceptedVersion

Denne oppgaven forsøker å finne ut hvem, om noen, som kan ha opphavsrett til bilder produsert av maskinlærings-modeller. Modeller som Stable Diffusion er verktøy med mange muligheter i stadig utvikling og oppgaven vil behandle verkshøyde-kravet, det grunnleggende kravet for opphavsrett, for å finne ut hvilke mengder og typer menneskelig innflytelse som kan gi opphavsrett til sluttproduktet. Videre vil oppgaven se på ulike bruksscenarioer for kunstig intelligens for å vurdere om de oppfyller verkshøyde-kravet eller ikke, og hva som er mulige konsekvenser av det. Som oppgaven vil vise er modellene avhengig av menneskelig assistanse for å produsere bilder, men det er også tenkelige tilfeller hvor programmet assisterer menneskelig skapelse heller enn det motsatte.

Recurrent neural networks (RNN), which are able to capture temporal natures of a signal, are becoming more common in machine learning applied to petroleum engineering, particularly drilling. With this technology come requirements and caveats related to the input data that play a significant role on resultant models. This paper explores how data pre-processing and attribute selection techniques affect the RNN models’ performance. Re-sampling and down-sampling methods are compared; imputation strategies, a problem generally omitted in published research, are explored and a method to select either last observation carried forward or linear interpolation is introduced and explored in terms of model accuracy. Case studies are performed on real-time drilling logs from the open Volve dataset published by Equinor. For a realistic evaluation, a semi-automated process is proposed for data preparation and model training and evaluation which employs a continuous learning approach for machine learning model updating, where the training dataset is being built continuously while the well is being made. This allows for accurate benchmarking of data pre-processing methods. Included is a previously developed and updated branched custom neural network architecture that includes both recurrent elements as well as row-wise regression elements. Source code for the implementation is published on GitHub.

For robotic technology to be adopted within the agricultural domain, there is a need for low-cost systems that can be flexibly deployed across a wide variety of crop types, environmental conditions, and planting methods, without extensive re-engineering. Here we present an approach for predicting steering angles for an autonomous, crop row-following, agri-robot using only RGB image input. Our approach employs a deep convolutional neural network (DCNN) and an end-to-end learning strategy. We pre-train our network using existing open datasets containing natural features and show that this approach can help to preserve performance across diverse agricultural settings. We also present preliminary results from open-loop field tests that demonstrate the feasibility and some of the limitations of this approach for agri-robot guidance.

PhD thesis in Information technology At present the influence of social media on society is so much that without it life seems to have no meaning for many. This kind of over-reliance on social media gives an opportunity to the anarchic elements to take undue advantage. Online misinformation and deception are vivid examples of such phenomenon. The misinformation or fake news spreads faster and wider than the true news [32]. The need of the hour is to identify and curb the spread of misinformation and misleading content automatically at the earliest. Several machine learning models have been proposed by the researchers to detect and prevent misinformation and deceptive content. However, these prior works suffer from some limitations: First, they either use feature engineering heavy methods or use intricate deep neural architectures, which are not so transparent in terms of their internal working and decision making. Second, they do not incorporate and learn the available auxiliary and latent cues and patterns, which can be very useful in forming the adequate context for the misinformation. Third, Most of the former methods perform poorly in early detection accuracy measures because of their reliance on features that are usually absent at the initial stage of news or social media posts on social networks. In this dissertation, we propose suitable deep neural attention based solutions to overcome these limitations. For instance, we propose a claim verification model, which learns embddings for the latent aspects such as author and subject of the claim and domain of the external evidence document. This enables the model to learn important additional context other than the textual content. In addition, we also propose an algorithm to extract evidential snippets out of external evidence documents, which serves as explanation of the model’s decisions. Next, we improve this model by using improved claim driven attention mechanism and also generate a topically diverse and non-redundant multi-document fact-checking summary for the claims, which helps to further interpret the model’s decision making. Subsequently, we introduce a novel method to learn influence and affinity relationships among the social media users present on the propagation paths of the news items. By modeling the complex influence relationship among the users, in addition to textual content, we learn the significant patterns pertaining to the diffusion of the news item on social network. The evaluation shows that the proposed model outperforms the other related methods in early detection performance with significant gains. Next, we propose a synthetic headline generation based headline incongruence detection model. Which uses a word-to-word mutual attention based deep semantic matching between original and synthetic news headline to detect incongruence. Further, we investigate and define a new task of incongruence detection in presence of important cardinal values in headline. For this new task, we propose a part-of-speech pattern driven attention based method, which learns requisite context for cardinal values.

Land cover maps are important tools for quantifying the human footprint on the environment and facilitate reporting and accounting to international agreements addressing the Sustainable Development Goals. Widely used European land cover maps such as CORINE (Coordination of Information on the Environment) are produced at medium spatial resolutions (100 m) and rely on diverse data with complex workflows requiring significant institutional capacity. We present a 10 m resolution land cover map (ELC10) of Europe based on a satellite-driven machine learning workflow that is annually updatable. A random forest classification model was trained on 70K groundtruth points from the LUCAS (Land Use/Cover Area Frame Survey) dataset. Within the Google Earth Engine cloud computing environment, the ELC10 map can be generated from approx. 700 TB of Sentinel imagery within approx. 4 days from a single research user account. The map achieved an overall accuracy of 90% across eight land cover classes and could account for statistical unit land cover proportions within 3.9% (R2 = 0.83) of the actual value. These accuracies are higher than that of CORINE (100 m) and other 10 m land cover maps including S2GLC and FROM-GLC10. Spectrotemporal metrics that capture the phenology of land cover classes were most important in producing high mapping accuracies. We found that the atmospheric correction of Sentinel-2 and the speckle filtering of Sentinel-1 imagery had a minimal effect on enhancing the classification accuracy (< 1%). However, combining optical and radar imagery increased accuracy by 3% compared to Sentinel-2 alone and by 10% compared to Sentinel-1 alone. The addition of auxiliary data (terrain, climate and night-time lights) increased accuracy by an additional 2%. By using the centroid pixels from the LUCAS Copernicus module polygons we increased accuracy by

Graphical abstract, Antibiotic resistance poses a major threat to public health. More effective ways of the antibiotic prescription are needed to delay the spread of antibiotic resistance. Employment of sequencing technologies coupled with the use of trained neural network algorithms for genotype-to-phenotype prediction will reduce the time needed for antibiotic susceptibility profile identification from days to hours. In this work, we have sequenced and phenotypically characterized 171 clinical isolates of Escherichia coli and Klebsiella pneumoniae from Norway and India. Based on the data, we have created neural networks to predict susceptibility for ampicillin, 3rd generation cephalosporins and carbapenems. All networks were trained on unassembled data, enabling prediction within minutes after the sequencing information becomes available. Moreover, they can be used both on Illumina and MinION generated data and do not require high genome coverage for phenotype prediction. We cross-checked our networks with previously published algorithms for genotype-to-phenotype prediction and their corresponding datasets. Besides, we also created an ensemble of networks trained on different datasets, which improved the cross-dataset prediction compared to a single network. Additionally, we have used data from direct sequencing of spiked blood cultures and found that AMR-Diag networks, coupled with MinION sequencing, can predict bacterial species, resistome, and phenotype as fast as 1–8 h from the sequencing start. To our knowledge, this is the first study for genotype-to-phenotype prediction: (1) employing a neural network method; (2) using data from more than one sequencing platform; and (3) utilizing sequence data from spiked blood cultures.

Hjertestans er den vanligste årsaken til plutselig død blant idrettsutøvere. Dette skyldes ofte arvelige hjertesykdommer som blant annet kan oppdages på et elektrokardiogram (EKG). I denne studien ble to problemstillinger undersøkt: (1) Om tolkningsalgoritmene i en klinisk anvendt elektrokardiograf er bedre tilpasset EKG fra den generelle befolkning enn idrettsutøvere? (2) Om moderne metoder, som kunstig intelligens (KI), kunne brukes til å forbedre dagens kliniske anvendte tolkningsalgoritmer slik at de kan bli bedre kalibrert for idrettsutøvere? For å undersøke (1) ble det registrert EKG fra 28 idrettsutøvere ved bruk av en moderne elektrokardiograf. EKGene ble tolket av tolkningsalgoritmen i elektrokardiografen, og sammenlignet med en kardiologs tolkning. Tolkningsalgoritmen klassifiserte 46% av EKGene som normale, 25% i grenseområdet og 29% som unormale. Kardiologen fulgte retningslinjene fra International Criteria, og klassifiserte 93% som normale og 7% i grenseområdet. Resultatene viser at tolkningsalgoritmen feilvurderer noen av områdene i EKGet. Dette ga i noen tilfeller alvorlige feilklassifiseringer, som akutt hjerteinfarkt (STEMI). For å undersøke (2) ble et stort åpent datasett av EKGer fra den generelle befolkning tatt i bruk. Ni ulike modeller ble trent, og validert på datasettet, deretter ble fire prediksjoner fra de to beste modellene forklart ved bruk av forklarbar KI. Den beste modellen fikk en PhysioNet/CinC Challenge skår = 0,512±0,006 og en F1-skår=0,405±0,008 ved 10-ganger kryssvalidering. Modellarkitekturene lar seg forklare ved bruk av forklarbar KI. Forklarbarheten er viktig for modellens kliniske relevans. For videre arbeid med disse problemstillingene trengs tilgang til flere EKG fra idrettsutøvere.

Bruken av maskinlæring reiser mange problemstillinger og utfordrer prinsippet om lovlighet, åpenhet og rettferdighet på flere måter. I oppgaven min har jeg gjennomført enkeltstudiet om skatteetatens praksis med maskinlæringsteknologi og samlet en del interessante funn. Med dette bidraget ønsker jeg å formidle utfordringene med utvikling av risikoscoringsmodellen som brukes i skattekontroll av mva-oppgaver. I samme anledningen har jeg redegjort for en del juridiske krav om transparens som stilles vedrørende profilering. Mot slutten av undersøkelsen har jeg kartlagt en del tiltak som bidrar til et godt personvern.

Denne oppgaven dreier seg om NAVs arbeid med å utvikle et beslutningsstøttesystem som baserer seg på maskinlæring. Systemet skal brukes i forbindelse med vurderingen av behovet for dialogmøte 2 i sykefraværsoppfølgingen. Utviklingsarbeidet jeg har undersøkt er ikke ferdigstilt, og oppgaven bør derfor leses som en utforskning og beskrivelse av en utviklingsprosess, ikke av et system i bruk. Forskningsspørsmålene i studien er eksplorerende, og forskningsdesignet er en case-studie. For å undersøke utviklingsarbeidet, tar jeg utgangspunkt i ideer fra critical data studies. Jeg har dermed undersøkt hvordan NAV jobber med å tilrettelegge og analysere datasett. I tillegg har jeg trukket linjer til hvordan organisering og rettslige vurderinger påvirker dette arbeidet. Et hovedpoeng har vært å synliggjøre utfordringer NAV møter på i utviklingsarbeidet. Utviklingsteamet har enn så lenge brukt mye tid på å kartlegge og klargjøre datagrunnlaget de kan bruke, og mindre på å utvikle algoritmen som skal analysere datagrunnlaget. Organisatorisk er det primært teknologer, data scientister, og veiledere i NAV som påvirker systeminnretningen. Enkelte eksterne virksomheter påvirker arbeidet, primært innenfor forklarbarhet. Forholdet til intern juridisk kompetanse er preget av at juristene i NAV har svært lite tid. Dette påvirker både omfanget av samarbeidet mellom teamet og jurister, men også de mange og prinsipielle vurderingene juristene må ta i et slikt arbeid. Regelverket som gjelder for bruk av maskinlæring i saksbehandlingen til offentlig sektor er svært uklart. Det er stor usikkerhet rundt tolkning av det gjeldende regelverket, blant annet for forklarbarhet, rettferdighet, dataminimering, og hjemmelsgrunnlaget. Det er utfordrende for enkeltjurister i NAV å skulle ta slike omfattende og prinsipielle vurderinger. Jeg konkluderer derfor med å peke på behovet for at lovgiver tar tydelig stilling til disse spørsmålene, istedenfor å overlate det til rettsanvenderne.

In the offshore industry, unmanned autonomous systems are expected to have a permanent role in future operations. During offshore operations, the unmanned autonomous system needs definite instructions on evaluating the gathered data to make decisions and react in real-time when the situation requires it. We rely on video surveillance and sensor measurements to recognize early warning signals of a failing asset during the autonomous operation. Missing out on the warning signals can lead to a catastrophic impact on the environment and a significant financial loss. This research is helping to solve the issue of trustworthiness of the algorithms that enable autonomy by capturing the rising risks when machine learning unintentionally fails. Previous studies demonstrate that understanding machine learning algorithms, finding patterns in anomalies, and calibrating trust can promote the system’s reliability. Existing approaches focus on improving the machine learning algorithms and understanding the shortcomings in the data collection. However, recollecting the data is often an expensive and extensive task. By transferring knowledge from multiple disciplines, diverse approaches will be observed to capture the risk and calibrate the trust in autonomous systems. This research proposes a conceptual framework that captures the known risks and creates a safety net around the autonomy-enabling algorithms to improve the reliability of the autonomous operations. Reliable Unmanned Autonomous Systems: Conceptual Framework for Warning Identification during Remote Operations

Condition-Based Monitoring (CBM) systems have grown in popularity in recent years owing to innovations in areas, such as sensor-technology, communication systems, and computing. That has fostered the development of more efficient systems to monitor, analyze, and identify failures in industrial plants, production lines, and machinery. Gas and oil industries lose billions of dollars yearly related to abnormal events and systems failures. Thus, Abnormal Event Management (AEM), which aims at early detection and identification of these events, has become their number one priority so that preventive actions can be taken timely. This work addresses the issue of detection and classification of faults in offshore oil wells. The aim is to create a CBM system based on the random forest classifier to support decision-making. The events used in this work are part of the 3W database developed by Petrobras, Brazil, one of the world's largest oil producer. Seven events categorized as faulty events are considered, as well as several instances labeled as normal operation. We conducted two experiments related to two different classification scenarios. The proposed systems achieved an overall accuracy of 90\%, indicating that the system is not only able to detect faulty events but can also anticipate incoming failures successfully.

Masteroppgave i innovasjon og ledelse, Høgskulen på Vestlandet, campus Bergen Kunstig intelligens og Big Data er digitale teknologier som blir stadig mer relevant i flere bransjer, også i forsikringsbransjen som vi studerer i denne oppgaven. Studien belyser hvordan forsikringsbransjen på tvers av konkurrerende selskaper kan ta i bruk KI og Big Data for å bekjempe forsikringssvindel, og hvilke utfordringer som kan oppstå ved en slik løsning. Vi har derfor valgt å studere et innovasjonsprosjekt som jobber med å validere et konsept for trygg datadeling mellom konkurrerende selskaper for å bekjempe et felles problem. Vi fokuserer på de teknologiske, forretningsmessige og juridiske dimensjonene ved prosjektet. Det teoretiske rammeverket bygger derfor på flere begrep; innovasjon, åpen innovasjon, samkonkurranse, kunstig intelligens og personvern/GDPR. Oppgaven er en kvalitativ single case studie hvor vi studerer ett spesifikt innovasjonsprosjekt. Vi har i den forbindelse samlet inn data ved hjelp av semi-strukturerte dybdeintervjuer og gruppeintervjuer med deltakere i prosjektet. Kvalitativ metode ble ansett som mest hensiktsmessig ettersom konseptet vi studerer er innenfor et relativt nytt felt hvor det eksisterer lite forskning fra før. Ved å snakke med alle involverte parter i prosjektet har vi innhentet ulike synspunkt, men også fått avdekket at deltakerne i prosjektet har en felles forståelse om hva som er deres største MOØ300

With an ever-increasing amount of image data, the manual labeling process has become the bottleneck in many machine learning applications. Plankton taxa labeling is especially a challenge due to its complex nature, and the manual labeling effort places a large burden on the domain experts. The Active Learning (AL) paradigm is a promising research direction adopted in the literature to minimize the manual labeling effort exerted by domain experts. Many approaches for AL have been proposed over the recent years to improve the labeling task by supporting the construction of large data sets suitable to train machine learning models while minimizing human involvement in the process. Our empirical study suggests that many modern active learning methods fail to incorporate both the samples that represent the statistical pattern of the data and the samples in which the machine learning model is not confident about. Inspired by these limitations, we propose an algorithm that combines these two types of sampling in order to capture the data distribution of the whole feature space, prevent redundant sampling from correlated uncertainty queries and fine-tune the inter-class decision boundary. Our experiments show that the proposed method outperforms each of the methods separately Further, it also proves to be efficient on both the CIFAR-10 data set and the more complex Kaggle plankton dataset.

Masteroppgave i innovasjon og ledelse, Høgskulen på Vestlandet, campus Bergen Denne empiriske studien undersøker barrierer og drivere i AI-realiseringsprosesser i virksomheter hvilket er i prosessen med å implementere kunstig intelligens for et kommersielt formål. Det var Alan Turing som introduserte ideen om at maskiner kunne tenke allerede i 1950, men det er først i det siste tiåret at kunstig intelligens har gått fra å være et rent teoretisk konsept til å bli en reell anvendelig teknologi. Studiens formål er derfor å identifisere barrierer og drivere som påvirker et AI-initiativ, samt forutsetninger, hensyn og kjennetegn ved slike initiativ. Studien undersøker fenomenet AI-initiativ gjennom en kvalitativ multippel casestudie med en kombinasjon av deduktiv og induktiv forskningstilnærming. Datagrunnlaget er innhentet gjennom ni semi-strukturere kvalitative dybdeintervjuer med ledere og teknologer fra studiens fem case-virksomheter, hvilket opererer i flere ulike bransjer, samt et ekspert-intervju med prosjektleder i Norwegian Cognitive Center. Studien har avdekket at nivå av Artificial Narrow Intelligence (ANI) kompleksitet vil være avgjørende for hvilke barrierer og drivere virksomheter støter på under AI-realiseringsprosessen. De tre identifiserte overordnede nivåene av ANI-kompleksitet er; (1) Integration of third-party AI APIs, (2) Modification of third-party AI APIs, and (3) Development of AI models, where level 3 is the top level of complexity. The level of complexity will be a determining factor for which barriers and facilitating factors companies face within the aspects of scope & complexity, organizational circumstances, knowledge & competence, and data. Barriers and facilitating factors associated with organizational flexibility, privacy regulations, and motivation have been found to be applicable regardless of ANI complexity. Prior to the data collection, we developed an AI realization model based on several different literary contributions, where the AI realization process was presented as a linear process. On the other hand, empirical findings indicated that the AI realization process was a continuous and iterative process, where AI initiatives will become part of the companies' daily operations over time if they succeed with the process. Motivation was also revealed as an important aspect in the AI realization process as an initial and continuous facilitating factor. Findings also indicate greater uncertainty and lower predictability associated with AI initiatives organized in projects, compared to other development projects. This uncertainty seems to be augmented by a lack of AI expertise. MOØ300

Image segmentation is one of the key components in systems performing computer vision recognition tasks. Various algorithms for image segmentation have been developed in the literature. Among them, more recently, deep learning algorithms have been remarkably successful in performing this task. A downside with deep neural networks for segmentation is that they require a large amount of labeled dataset for training. This prerequisite is one of the main reasons that led researchers to adopt data augmentation approaches in order to minimize manual labeling efforts while maintaining highly accurate results. This paper uses classical non-deep learning methods for background extraction to increase the size of the dataset used to train deep learning attention segmentation algorithms when images are presented as time-series to the model. The method presented adopts the Gaussian mixture-based (MOG2) foreground-background segmentation followed by dilation and erosion to create masks necessary to train the deep learning models. It is applied in the context of planktonic images captured in situ as time series. Various evaluation metrics and visual inspection are used to compare the performance of the deep learning algorithms. Experimental results show higher accuracy achieved by the deep learning algorithms for time-series image attention segmentation when the proposed data augmentation methodology is utilized to increase the training dataset.

Background The Living Evidence Map Project at the Norwegian Institute of Public Health (NIPH) gives an updated overview of research results and publications. As part of NIPH’s mandate to inform evidence-based infection prevention, control and treatment, a large group of experts are continously monitoring, assessing, coding and summarising new COVID-19 publications. Screening tools, coding practice and workflow are incrementally improved, but remain largely manual. Results This paper describes how deep learning methods have been employed to learn classification and coding from the steadily growing NIPH COVID-19 dashboard data, so as to aid manual classification, screening and preprocessing of the rapidly growing influx of new papers on the subject. Our main objective is to make manual screening scalable through semi-automation, while ensuring high-quality Evidence Map content. Conclusions We report early results on classifying publication topic and type from titles and abstracts, showing that even simple neural network architectures and text representations can yield acceptable performance.

Model errors are ubiquitous in practical history matching problems. A common approach in the literature to accounting for model errors is to treat them as random variables following certain presumed distributions. While such a treatment renders algorithmic convenience, its underpinning assumptions are often invalid. In this work, we adopt an alternative approach, and treat model-error characterization as a functional approximation problem, which can be solved using a generic machine learning method. We then integrate the proposed model- error characterization approach into an ensemble-based history matching framework, and show that, with very minor modifications, existing ensemble-based history matching algorithms can be readily deployed to solve the history matching problem in the presence of model errors. To demonstrate the efficacy of the integrated history matching framework, we apply it to account for potential model errors of a rock physics model in 4D seismic history matching applied to the full Norne benchmark case. Our experiment results indicate that the proposed model-error characterization approach helps improve the qualities of estimated reservoir models, and leads to more accurate forecasts of production data. This suggests that accounting for model errors from a perspective of machine learning may serve as a novel and viable way to deal with model imperfection in practical history matching problems.

In recent years, there were multiple papers published related to rate of penetration prediction using machine learning vastly outperforming analytical methods. There are models proposed reportedly achieving R2 values as high as 0.996. Unfortunately, it is most often impossible to independently verify these claims as the input data is rarely accessible to others. To solve this problem, this paper presents a database derived from Equinor's public Volve dataset that will serve as a benchmark for rate of penetration prediction methods. By providing a partially processed dataset with unambiguous testing scenarios, scientists can perform machine learning research on a level playing field. This in turn will both discourage publication of methods tested in a substandard manner as well as promote exploration of truly superior solutions. A set of seven wells with nearly 200–000 samples and twelve common attributes is proposed together with reference results from common machine learning algorithms. Data and relevant source code are published on the pages of University of Stavanger and GitHub.

Feature selection is an essential step in data science pipelines to reduce the complexity associated with large datasets. While much research on this topic focuses on optimizing predictive performance, few studies investigate stability in the context of the feature selection process. In this study, we present the Repeated Elastic Net Technique (RENT) for Feature Selection. RENT uses an ensemble of generalized linear models with elastic net regularization, each trained on distinct subsets of the training data. The feature selection is based on three criteria evaluating the weight distributions of features across all elementary models. This fact leads to the selection of features with high stability that improve the robustness of the final model. Furthermore, unlike established feature selectors, RENT provides valuable information for model interpretation concerning the identification of objects in the data that are difficult to predict during training. In our experiments, we benchmark RENT against six established feature selectors on eight multivariate datasets for binary classification and regression. In the experimental comparison, RENT shows a well-balanced trade-off between predictive performance and stability. Finally, we underline the additional interpretational value of RENT with an exploratory post-hoc analysis of a healthcare dataset.

This thesis identifies various explanations for attitudes towards climate change. My argument is that there are several explanations in the literature on perceptions of climate change that are not elucidated. These explanations are fragmented. The methods used in the existing literature may limit the number of possible explanatory factors. Moreover, the methods and the number of explanatory factors in the previous literature can cause the models to be underspecified. This may lead to important control variables being excluded from the analysis. In addition the methods are not well suited to capture non-linear relationships and interaction effects. In sum this makes it difficult to rely on the results from the previous literature. The thesis puts together theoretical threads from the existing literature, as well as adding obvious but neglected threads from the established voting literature. This creates a high number of potential (new and old) explanatory variables, which is the basis of machine learning as a method: a suitable empirical tool for feature selection. Putting this argument to test, the thesis uses machine learning with data from European Social Survey from 23 countries. The results from the analysis show that important explanatory variables from the traditional voting literature are crucial for predicting attitudes towards climate change. The group of important explanatory variables from the analysis can be divided into social controls, trust/satisfaction and values. The findings are not unambiguous but indicate that the classic explanatory factors that are established and widely tested in political science also apply to a new field of study. The voting literature is often important to investigate other issues and for the establishment of new theories in political science, such as post-material theories. Consequently the results is not surprising. Masteroppgave SAMPOL350 MASV-SAPO

Master's thesis in Petroleum Geosciences Engineering As petroleum geosciences enter the era of big data, this field of study encompass difficult optimization and decision-making in real-world problems. The increasing number, difficulty, and scale of such specific problems has become too complicated for geoscientists to rely on a single discipline for effective solution. Machine Learning (ML) provides extensive capability to be the solution in this area. This thesis, as a part of ongoing research, focuses on the application of various ML algorithms in predicting the net to gross value of the Brent Group reservoir zone in the Statfjord Field. For this purpose, several objectives were defined. As the first fundamental step, features were generated directly from the TWT and amplitudes of the respective seismic horizons. Secondly, predictive models were built from both training and testing phase using the features. The final task was estimating and mapping the value of net to gross property of the pre-defined reservoir zone. Furthermore, classification task and sand thickness prediction were also included as additional comparisons to the main task. The results indicate outstanding performance demonstrated by Decision Tree and Random Forest algorithms despite the limitation on the dataset. Insufficient amount of data as well as data cleaning problems have been the main constraints in this study. This unarguably led to high variance in the data which yielded less accurate and less reliable prediction models. The ML clearly have potential to accomplish the defined task better once the obstacles are handled properly in the future studies. Some improvements such as better data cleaning process, more involvement of well logs data, AVO inversion analysis, and utilization of more advance algorithms are strongly suggested in order to boost the models’ performance.

Master's thesis in Industrial economics Today's industrial era is rapidly changing due to digitization of information and automation of work processes. Digitalization is disrupting several industries and the vision of fully autonomous production facilities, also known as Industry 4.0, is becoming more and more likely every day. In all manufacturing and production facilities maintenance is a core activity to secure reliability, availability, and safety. This is a wildly accepted truth in asset heavy industries and digitalization projects in multiple industries have been started to take advantage of available technologies. For the oil and gas (O&G) industry, digitalization has focused on making assets smart through sensing technology and digital control systems. On the Norwegian Continental Shelf offshore oil and gas platforms are well equipped with condition monitoring equipment that collect performance- and health data continuously. The problem arises when analyzing aggregated data as decision support for maintenance planning. Trending and subjective opinions are still the most used methodologies even though artificial intelligence (AI) and Big Data analytics could improve this task significantly. AI has the potential to screen huge quantities of data and discover patterns, correlations and calculate statistical possibilities which would improve fault detectability, fault recognition (i.e. diagnostics) and remaining useful life and degradation rate (i.e. prognostics) accuracy. To implement predictive AI analytics there are a set of predetermined requirements that need to be included in the maintenance work process. However, the state-of-practice self-assessment methodology (i.e. Maintenance Baseline Study) does not include state-of-the-art technologies such as AI and ML analytic tools, nor maintenance strategies as predictive maintenance. Thus, the purpose of this thesis is to develop a new self-assessment model where readiness to implement AI analytics is the governing theme. With the Analyses Module in the Maintenance Baseline Study as a benchmark, the goal was to develop an easy to use, accessible and up-to-date self-assessment model that allows companies to assess their readiness to implement AI analytics tools related to a predictive maintenance strategy. The updated self-assessment model was developed after a thorough literature review of operation- and maintenance theory and Industry 4.0 enabling technologies. This theoretical background formed the basis of the new model, and further developments helped achieve the following improvements: 1. An online self-assessment platform provides accessibility from any device connected to the internet. 2. Multiple-choice reduces complexity, assessment time and provides all responses in a quantitative format which simplifies group / industry analysis if desirable. 3. The models focuses on predictive maintenance and AI technologies are visionary, i.e. the insight received by conducting the self-assessment is essential information for building a system for predictive analysis. An important part of model development was review and feedback gathering from operation- and maintenance professionals. Through an iterative process of review, feedback and updates, the model was highly influenced by industrial expertise. This was an important measure to secure its validation, and to verify its value.

Short-term load forecasting ensures the efficient operation of power systems besides affording continuous power supply for energy consumers. Smart meters that are capable of providing detailed information on buildings energy consumption, open several doors of opportunity to short-term load forecasting at the individual building level. In the current paper, four machine learning methods have been employed to forecast the daily peak and hourly energy consumption of domestic buildings. The utilized models depend merely on buildings historical energy consumption and are evaluated on the profiles that were not previously trained on. It is evident that developing data-driven models lacking external information such as weather and building data are of great importance under the situations that the access to such information is limited or the computational procedures are costly. Moreover, the performance evaluation of the models on separated house profiles determines their generalization ability for unseen consumption profiles. The conducted experiments on the smart meter data of several UK houses demonstrated that if the models are fed with sufficient historical data, they can be generalized to a satisfactory level and produce quite accurate results even if they only use past consumption values as the predictor variables. Furthermore, among the four applied models, the ones based on deep learning and ensemble techniques, display better performance in predicting daily peak load consumption than those of others.

The timeliness of detection of a sepsis event in progress is a crucial factor in the outcome for the patient. Machine learning models built from data in electronic health records can be used as an effective tool for improving this timeliness, but so far the potential for clinical implementations has been largely limited to studies in intensive care units. This study will employ a richer data set that will expand the applicability of these models beyond intensive care units. Furthermore, we will circumvent several important limitations that have been found in the literature: 1) Models are evaluated shortly before sepsis onset without considering interventions already initiated. 2) Machine learning models are built on a restricted set of clinical parameters, which are not necessarily measured in all departments. 3) Model performance is limited by current knowledge of sepsis, as feature interactions and time dependencies are hardcoded into the model. In this study, we present a model to overcome these shortcomings using a deep learning approach on a diverse multicenter data set. We used retrospective data from multiple Danish hospitals over a seven-year period. Our sepsis detection system is constructed as a combination of a convolutional neural network and a long short-term memory network. We suggest a retrospective assessment of interventions by looking at intravenous antibiotics and blood cultures preceding the prediction time. Results show performance ranging from AUROC 0.856 (3 hours before sepsis onset) to AUROC 0.756 (24 hours before sepsis onset). We present a deep learning system for early detection of sepsis that is able to learn characteristics of the key factors and interactions from the raw event sequence data itself, without relying on a labor-intensive feature extraction work.

The Swedish procurement process has a yearly value of 706 Billion SEK over approximately 18 000 procurements. With each process comes many documents written in different formats that need to be understood to be able to be a possible tender. With the development of new technology and the age of Machine Learning it is of huge interest to investigate how we can use this knowledge to enhance the way we procure. The goal of this project was to investigate if public procurements written in Swedish in PDF format can be parsed and segmented into a structured format. This process was divided into three parts; pre-processing, annotation, and training/evaluation. The pre-processing was accomplished using an open-source pdf-parser called pdfalto that produces structured XML-files with layout and lexical information. The annotation process consisted of generalizing a procurement into high-level segments that are applicable to different document structures as well as finding relevant features. This was accomplished by identifying frequent document formats so that many documents could be annotated using deterministic rules. Finally, a linear chain Conditional Random Field was trained and tested to segment the documents. The models showed a high performance when they were tested on documents of the same format as it was trained on. However, the data from five different documents were not sufficient or general enough to make the model able to make reliable predictions on a sixth format that it had not seen before. The best result was a total accuracy of 90,6% where two of the labels had a f1-score above 95% and the two other labels had a f1-score of 51,8% and 63,3%.

Denne masteroppgaven er en casestudie som følger IKEA gjennom et pilotprosjekt om implementering av chatbot i kundeservice våren 2020. Studien deles inn i tre stadier, hvor første stadiet handler om identifisering av behovet for en chatbot i IKEA. Det andre stadiet handler om konstruksjon av chatboten og redegjør for hvilke muligheter og begrensninger som eksisterer i IKEA for implementering. Etter-stadiet fokuserer på de initiale effektene av chatboten etter implementering, samt hvilke faktorer som bidrar til at kunder av IKEA benytter chatboten. Hele implementeringsprosessen redegjøres på bakgrunn av både kvalitativt og kvantitativt datagrunnlag, som består av intervjuer med IKEA og de involverte partene Boost Ai og NetNordic. For å få bredere innsikt omkring chatbot og andre bransjer, ble også Kantar intervjuet. En forskningsmodell ble også utarbeidet på grunnlag av en spørreundersøkelse rettet mot kunder av IKEA. Resultatet fra første stadiet viser at endrede kundeforventninger, økt kontaktvolum og kompetanseheving viser behovet av en chatbot i IKEAs kundeservice. Resultatene i det andre stadiet viser at en chatbot har mulighet til å besvare henvendelser som relateres til FAQ og ordresporing. IKEA har også mulighet til å sette sammen en prosjektgruppe bestående av ulik kompetanse og erfaringer fra ulike områder. IKEA har begrensninger knyttet til dårlig koordinasjon med den globaliserte IT-avdelingen, eldre IT-systemer og mangler på APIer. Siste stadiet viser at chatboten har en begrenset effekt i reduksjon av kontaktvolumet fordi mange av kundehenvendelsene faller utenfor dens område. Det eksisterer likevel positive effekter ved at chatboten skaper verdi for de kundehenvendelsene som predikeres riktig, samt at den predikerer store mengder med kundehenvendelser samtidig. I tillegg opererer den døgnkontinuerlig. Skalerbarheten i chatboten gjør at den tilpasser seg endrede omgivelser relativt fort. Resultatene fra forskningsmodellen viser at løsningsgraden til chatboten må være høy, operere brukervennlig og fascinere for at kunder av IKEA vil bruke chatboten. Vår masteroppgave viser at detaljhandel er preget av komplekse vare- og tjenesteporteføljer, som skaper utfordringer for en begrenset kognitiv chatbot i kundeservice. Dette tyder på at kunstig intelligens har en vei å gå før den ønskede merverdien blir oppfylt. This master’s thesis is a case study that seeks to analyze the implementation process of a chatbot in IKEA’s customer service. The thesis follows IKEA through a pilot project during the spring of 2020. The first stage involves the identification of the need of a chatbot in IKEA’s customer service. The second stage involves the construction of the chatbot and clarifies the possibilities and limitations that exist in IKEA in relation to the implementation. The after stage focuses on the initial effects after the implementation and what factors contribute to customers of IKEA using the chatbot. The implementation process is explained on the basis of both qualitative and quantitative data, which consist of interviews with IKEA, Boost Ai and NetNordic. In order to gain a broader understanding of chatbot and other industries, we also interviewed Kantar. A research model was also developed based on a survey aimed at customers of IKEA. Our results in the first stage show that changes in customers’ expectations, an increase in contact volume and competence development indicates to a need for a chatbot in IKEA’s customer service. The second stage shows that the chatbot has the possibility to answer inquiries related to FAQ and order tracking. IKEA also has the possibility to assemble a project team, consisting of mixed competence and experience. IKEA has limitations related to poor coordination with it’s global IT-department, older IT-systems and lack of relevant APIs. The results of the after stage show that the chatbot creates a limited effect in reducing the contact volume because many of the customer inquiries fall outside of the chatbot’s limited domain. However, there are some positive effects related to correctly predicted inquiries, and the possibility to frequently predict customer inquiries at the same time. In addition, the chatbot operates 24 hours a day. The scalability of the chatbot allows it to quickly adapt to changing surroundings. The results of the research model show that performance expectancy, effort expectancy and hedonic motivation are important factors that explain the customers’ behavioral intention to use the IKEA chatbot. Our master’s thesis shows that the retail industry is characterized by complex product and service portfolios, which create challenges for a limited cognitive chatbot in customer service. This implies that artificial intelligence still has a long way to go before a chatbot is viable in the retail industry. M-ØA

Masteroppgave i innovasjon og ledelse, Høgskulen på Vestlandet, campus Bergen Da vi startet prosessen med å skrive denne masteroppgaven avdekket vi en unik mulighet for å utvikle en svært spennende oppgave ulik alle andres – med stor nyhetsverdi. Denne eksplorative multiple case-studien tar for seg hvordan innovasjonsselskaper i medieklyngen i Bergen forholder seg til kunstig intelligens. På grunn av fenomenets kompleksitet og mangel på tidligere forskning gikk vi bredt ut for å forsikre oss om at vi fikk belyst og dekket relevante områder. Gjennom denne iterative prosessen avdekket vi fire særlig interessante elementer i selskapene sine forretningsmodeller; verdiforslag, inntektsstrømmer, nøkkelaktiviteter og nøkkelpartnere. Videre så vi at det var relevant å undersøke dette i lys av selskapenes omgivelser. Vi valgte derfor å også fokusere på teknologiske omgivelser fra PESTEL-rammeverket, samt søylene i de institusjonelle omgivelsene. Målet har vært å avdekke hvordan disse innovasjonsselskapene forholder seg til kunstig intelligens og hva de vektlegger i omgivelsene, samt hvordan de håndterer dette, i tillegg til å belyse hvordan kunstig intelligens benyttes i selskapenes forretningsmodeller. Gjennom en rekke intervjuer med følgende case-selskaper; IBM, TV 2, Vimond og Mjoll, samt intervjuer med NCE Media, Departementenes sikkerhets- og serviceorganisasjon og Beat Technology, avdekket vi at krefter i de teknologiske omgivelsene presser på for endringer – samtidig som at krefter i de institusjonelle omgivelsene bremser graden av adopsjon av innovativ teknologi som kunstig intelligens. Videre fant vi at dette dilemmaet kan håndteres gjennom forretningsmodellinnovasjon. Samtlige case-selskaper benytter seg av kunstig intelligens av typen maskinlæring, til tross for at maskinlæringen på langt nær er optimal per i dag. Den forbedres imidlertid hurtig. Det er avgjørende at bruk av slik teknologi ikke skaper demokratiske utfordringer og at den håndteres i tråd med normer og verdier som tillit og troverdighet samt regelverk som GDPR. Selskapene legger til rette for økt forskning og utvikling samt adopsjon av maskinlæring gjennom FMI ved økt fokus på fleksibilitet og forbedringer i eksempelvis aktiviteter, samarbeid og prosjekter. MOØ300

Denne oppgaven er en case-studie i samarbeid med Frende Forsikring, og ser på bruken av maskinlæring som virkemiddel for å forbedre tilnærmingen til kundebevaring. Forskning viser til en betydelig høyere kostnad ved anskaffelse av nye kunder kontra bevaring av gamle, og en reduksjon i avgangsprosenten er derfor ønskelig. Problemet kundeavgang er i oppgaven delt inn i flere deler. Med dette er det skapt tre målvariabler, henholdsvis avgang, årsak til avgang og om kunde kan reddes, referert til som prediksjon 1,2 og 3. Utgangspunktet for oppdelingen er at det i realiteten kun er avgangskunder som avgår til nytt selskap det vil være mulig å redde. Totalt er syv forskjellige maskinlæringsmodeller blitt testet, hvorav de tre med best resultater ble fokus for videre arbeid. Fra de tre, alle basert på gradient boosting, er Light Gradient Boosted Machine (LightGBM) valgt som endelig modell for samtlige prediksjoner, og følgelig modellen analysen er basert på. For prediksjon 1 og 2 ble F1-score det viktigste målet å vurdere modellen etter, mens det for prediksjon 3 hovedsakelig ble vektlagt en kostnadsmatrise for valg av modell. Tolkning av modellens beslutningsprosess ble gjort ved å benytte SHAP-verdier, et mål som forklarer variablenes bidrag på utfallet, og ga grunnlag for gruppering av utsatte kunder. Resultatene viser at maskinlæring kan forbedre prosessen rundt kundebevaring. Potensielle avgangskunder kan bli identifisert på et tidlig stadium og arbeid med å redde kunder kan bli gjort med større presisjon. Dette resulterer i høyere verdi per redningsforsøk og dermed økt forretningsverdi. The thesis is a case study in collaboration with the Norwegian insurance company Frende and will be exploring the use of machine learning as a tool to improve the approach regarding customer retention. Research indicates a significantly greater cost of acquiring new customers compared to retaining old ones, and a reduction of the churn rate is therefore desirable. The problem of customer churn is divided into three categories. With this, three new target variables have been created, respectively churn, reason for churning and whether the customer can be saved. For future reference these are referred to as prediction 1, 2 and 3. The reason for this division, is that only customers who are leaving for a new insurance company are the ones it will be possible to save. M-ØA

Avhandlingen handler om opphavsrettslig vern av maskinlæringsalgoritmer, og opphavsrettslig vern til verk som genereres av maskinlæringsalgoritmer.