Insight Centre for Data Analytics - Doctoral Theses

Permanent URI for this collection


Recent Submissions

Now showing 1 - 5 of 22
  • Item
    Autonomous system control in unknown operating conditions
    (University College Cork, 2021-08-24) Sohège, Yves; Provan, Gregory; Tabirca, Marius-Sabin; Science Foundation Ireland
    Autonomous systems have become an interconnected part of everyday life with the recent increases in computational power available for both onboard computers and offline data processing. The race by car manufacturers for level 5 (full) autonomy in self-driving cars is well underway and new flying taxi service startups are emerging every week, attracting billions in investments. Two main research communities, Optimal Control and Reinforcement Learning stand out in the field of autonomous systems, each with a vastly different perspective on the control problem. Controllers from the optimal control community are based on models and can be rigorously analyzed to ensure the stability of the system is maintained under certain operating conditions. Learning-based control strategies are often referred to as model-free and typically involve training a neural network to generate the required control actions through direct interactions with the system. This greatly reduces the design effort required to control complex systems. One common problem both learning- and model- based control solutions face is the dependency on a priori knowledge about the system and operating conditions such as possible internal component failures and external environmental disturbances. It is not possible to consider every possible operating scenario an autonomous system can encounter in the real world at design time. Models and simulators are approximations of reality and can only be created for known operating conditions. Autonomous system control in unknown operating conditions, where no a priori knowledge exists, is still an open problem for both communities and no control methods currently exist for such situations. Multiple model adaptive control is a modular control framework that divides the control problem into supervisory and low-level control, which allows for the combination of existing learning- and model-based control methods to overcome the disadvantages of using only one of these. The contributions of this thesis consist of five novel supervisory control architectures, which have been empirically shown to improve a system’s robustness to unknown operating conditions, and a novel low- level controller tuning algorithm that can reduce the number of required controllers compared to traditional tuning approaches. The presented methods apply to any autonomous system that can be controlled using model-based controllers and can be integrated alongside existing fault-tolerant control systems to improve robustness to unknown operating conditions. This impacts autonomous system designers by providing novel control mechanisms to improve a system’s robustness to unknown operating conditions.
  • Item
    Improving human movement sensing with micro models and domain knowledge
    (2021-06-18) Scheurer, Sebastian; Brown, Kenneth; O'Sullivan, Barry; Science Foundation Ireland; European Regional Development Fund; Enterprise Ireland
    Human sensing is concerned with techniques for inferring information about humans from various sensing modalities. Examples of human sensing applications include human activity (or action) recognition, emotion recognition, tracking and localisation, identification, presence and motion detection, occupancy estimation, gesture recognition, and breath rate estimation. The first question addressed in this thesis is whether micro or macro models are a better design choice for human sensing systems. Micro models are models exclusively trained with data from a single entity, such as a Wi-Fi link, user, or other identifiable data-generating component. We consider micro and macro models in two human sensing applications, viz. Human Activity Recognition (HAR) from wearable inertial sensor data and device-free human presence detection from Wi-Fi signal data. The HAR literature is dominated by person-independent macro models. The few empirical studies that consider both micro and macro models evaluate them with either only one data-set or only one HAR algorithm, and report contradictory results. The device-free sensing literature is dominated by link-specific micro models, and the few papers that do use macro models do not evaluate their micro counterparts. Given the little and contradictory evidence, it remains an open question whether micro or macro models are a better design choice. We evaluate person-specific micro and person-independent macro models across seven HAR benchmark data-sets and four learning algorithms. We show that person-specific models (PSMs) significantly outperform the corresponding person-independent model (PIM) when evaluated with known users. To apply PSMs to data from new users, we propose ensembles of PSMs, which are improved by weighting their constituent PSMs according to their performance on other training users. We propose link-specific micro models to detect human presence from ambient Wi-Fi signal data. We select a link-specific model from the available training links, and show that this approach outperforms multi-link macro models. The second question addressed in this thesis is whether human sensing methods can be improved with domain knowledge. Specifically, we propose expert hierarchies (EHs) as an intuitive way to encode domain knowledge and simplify multi-class HAR, without negatively affecting predictive performance. The advantages of EHs are that they have lower time complexity than domain-agnostic methods and that their constituent classifiers are statistically independent. This property enables targeted tuning, and modular and iterative development of increasingly fine-grained HAR. Although this has inspired several uses of domain-specific hierarchical classification for HAR applications, these have been ad-hoc and without comparison to standard domain-agnostic methods. Therefore, it remains unclear whether they carry a penalty on predictive performance. We design five EHs and compare them to the best-known domain-agnostic methods. Our results show that EHs indeed can compete with more popular multi-class classification methods, both on the original multi-class problem and on the EHs' topmost levels.
  • Item
    Real-time algorithm configuration
    (University College Cork, 2021-04) Fitzgerald, Tadhg; O'Sullivan, Barry; Brown, Kenneth; Science Foundation Ireland; European Regional Development Fund
    This dissertation presents a number of contributions to the field of algorithm configur- ation. In particular, we present an extension to the algorithm configuration problem, real-time algorithm configuration, where configuration occurs online on a stream of instances, without the need for prior training, and problem solutions are returned in the shortest time possible. We propose a framework for solving the real-time algorithm configuration problem, ReACT. With ReACT we demonstrate that by using the parallel computing architectures, commonplace in many systems today, and a robust aggregate ranking system, configuration can occur without any impact on performance from the perspective of the user. This is achieved by means of a racing procedure. We show two concrete instantiations of the framework, and show them to be on a par with or even exceed the state-of-the-art in offline algorithm configuration using empirical evaluations on a range of combinatorial problems from the literature. We discuss, assess, and provide justification for each of the components used in our framework instantiations. Specifically, we show that the TrueSkill ranking system commonly used to rank players’ skill in multiplayer games can be used to accurately es- timate the quality of an algorithm’s configuration using only censored results from races between algorithm configurations. We confirm that the order that problem instances arrive in influences the configuration performance and that the optimal selection of configurations to participate in races is dependent on the distribution of the incoming in- stance stream. We outline how to maintain a pool of quality configurations by removing underperforming configurations, and techniques to generate replacement configurations with minimal computational overhead. Finally, we show that the configuration space can be reduced using feature selection techniques from the machine learning literature, and that doing so can provide a boost in configuration performance.
  • Item
    Asynchronous distributed clustering algorithms for wireless mesh network
    (University College Cork, 2021-04-10) Qiao, Cheng; Brown, Kenneth; O'Sullivan, Barry; Science Foundation Ireland; European Regional Development Fund
    Wireless Mesh Networks are becoming increasingly important in many applications. In many cases, data is acquired by devices that are distributed in space, but effective actions require a global view of that data across all devices. Transmitting all data to the centre allows strong data analytics algorithms to be applied, but consumes battery power for the nodes, and may cause data overload. To avoid this, distributed methods try to learn within the network, allowing each agent to learn a global picture and take appropriate actions. For distributed clustering in particular, existing methods share simple cluster descriptions until the network stabilises. The approaches tend to require either synchronous behaviour or many cycles, and omit important information about the clusters. In this thesis, we develop asynchronous methods that share richer cluster models, and we show that they are more effective in learning the global data patterns. Our underlying method describes the shape and density of each cluster, as well as its centroid and size. We combine cluster models by re-sampling from received models, and then re-clustering the new data sets. We then extend the approach, to allowing clustering methods that do not require the final number of clusters as input. After that, we consider the cases that there might be sub-groups of agents that are receiving different patterns of data. Finally, we extend our approaches to scenarios where each agent has no idea about whether there is a single pattern or are multiple patterns. We demonstrate that the approaches can regenerate clusters that are similar to the distributions that were used to create the test data. When the number of clusters are not available, the learned number of clusters is close to the ground truth. The proposed algorithms can learn how data points are clustered even when there are multiple patterns in the network. When the number of patterns (single or multiple) is not known in advance, the proposed methods Optimised KDE and DBSCAN preform well in detecting multiple patterns. Although they perform worse in detecting the single pattern, they can still learn how data points are clustered.
  • Item
    Active learning in recommender systems: an unbiased and beyond-accuracy perspective
    (University College Cork, 2020-12-05) Carraro, Diego; Bridge, Derek G.; O'Sullivan, Barry; Science Foundation Ireland; European Regional Development Fund
    The items that a Recommender System (RS) suggests to its users are typically ones that it thinks the user will like and want to consume. An RS that is good at its job is of interest not only to its customers but also to service providers, so they can secure long-term customers and increase revenue. Thus, there is a challenge in building better recommender systems. One way to build a better RS is to improve the quality of the data on which the RS model is trained. An RS can use Active Learning (AL) to proactively acquire such data, with the goal of improving its model. The idea of AL for RS is to explicitly query the users, asking them to rate items which have not been rated yet. The items that a user will be asked to rate are known as the query items. Query items are different from recommendations. For example, the former may be items that the AL strategy predicts the user has already consumed, whereas the latter are ones that the RS predicts the user will like. In AL, query items are selected `intelligently' by an Active Learning strategy. Different AL strategies take different approaches to identify the query items. As with the evaluation of RSs, preliminary evaluation of AL strategies must be done offline. An offline evaluation can help to narrow the number of promising strategies that need to be evaluated in subsequent costly user trials and online experiments. Where the literature describes the offline evaluation of AL, the evaluation is typically quite narrow and incomplete: mostly, the focus is cold-start users; the impact of newly-acquired ratings on recommendation quality is usually measured only for those users who supplied those ratings; and impact is measured in terms of prediction accuracy or recommendation relevance. Furthermore, the traditional AL evaluation does not take into account the bias problem. As brought to light by recent RS literature, this is a problem that affects the offline evaluation of RS; it arises when a biased dataset is used to perform the evaluation. We argue that it is a problem that affects offline evaluation of AL strategies too. The main focus of this dissertation is on the design and evaluation of AL strategies for RSs. We first design novel methods (designated WTD and WTD_H) that `intervene' on a biased dataset to generate a new dataset with unbiased-like properties. Compared to the most similar approach proposed in the literature, we give empirical evidence, using two publicly-available datasets, that WTD and WTD_H are more effective at debiasing the evaluation of different recommender system models. We then propose a new framework for offline evaluation of AL for RS, which we believe facilitates a more authentic picture of the performances of the AL strategies under evaluation. In particular, our framework uses WTD or WTD_H to mitigate the bias, but it also assesses the impact of AL in a more comprehensive way than the traditional evaluation used in the literature. Our framework is more comprehensive in at least two ways. First, it segments users in more ways than is conventional and analyses the impact of AL on the different segments. Second, in the same way that RS evaluation has changed from a narrow focus on prediction accuracy and recommendation relevance to a wider consideration of so-called `beyond-accuracy' criteria (such as diversity, serendipity and novelty), our framework extends the evaluation of AL strategies to also cover `beyond-accuracy' criteria. Experimental results on two datasets show the effectiveness of our new framework. Finally, we propose some new AL strategies of our own. In particular, our new AL strategies, instead of focusing exclusively on prediction accuracy and recommendation relevance, are designed to also enhance `beyond-accuracy' criteria. We evaluate the new strategies using our more comprehensive evaluation framework.