Abstract :
[en] Ambient Intelligence (AmI) constitutes a new paradigm of interaction among humans, smart objects and devices. AmI systems are expected to support humans in their every day tasks and activities. In order to achieve this goal, these systems require augmenting the environment with sensing, computing, communicating, and reasoning capabilities. Due to advances in technology, sensors are getting more powerful, cheaper and smaller, which stimulated large scale development and production. These sensors will generate a big amount of data and can easily lead to millions of values in a short amount of time, which can quickly reach the computation and storage limits when it comes to structuring and processing the data. For this problem, we propose a concept of continuous models that can handle highly-volatile data, and represent the continuous nature of sensor data in an efficient and compact way. We show on various AmI datasets that this can significantly improve storage and efficiency.
One important goal of AmI systems is to transform living and working environments into intelligent spaces able to adapt to their users’ needs and desires in real-time. In this sense, we call AmI applications context-aware, meaning that they use environmental information to adaptively provide more relevant and better services to the user. However, AmI systems are composed from heterogeneous components, operating in an open and dynamic environment. Each of these components can have different storage and computation capabilities. They might not have all the information needed to derive context information, and they might not be reachable all the time for various reasons. In this thesis, we present a contextual reasoning solution adapted for component based platforms. Our solution can derive contextual information in a distributed way and can handle inconsistencies when contradictory information is received from several components.
Other than the storage and computation efficiency, several qualities need to be satisfied according to the different contexts. Privacy is one of these qualities. AmI services will rely more and more on personal data that is vastly collected, stored, and exchanged with other third parties in order to provide added-value services. Such data are sensitive and often related to personal activities and therefore can lead to privacy risks, especially when data is shared with high precision and frequency. However, this privacy quality can be relaxed in some contexts, for example in an emergency situation in order to increase utility or efficiency. This leads to the need of developing an adaptive solution that is able to react to context changes in real-time and involve optimizing conflicting objectives. For this challenge, we propose to use blurring components as our main privacy preservation elements. The idea behind this approach is that, by gradually decreasing the data quality, a blurring component is able to hide some of the personal data delivered by sensors while still keeping the necessary information for the services to work. In order to find a good trade-off between these different conflicting objectives, we adapt a multi-objective evolutionary algorithm to run directly on top of domain specific models. We then apply it as our main optimization engine on models@run.time to keep adapting the different qualities, when the context change.
Finally, AmI services are expected to be tailored for different users’ needs in a seamless and unobtrusive way. Meaning that they should be able to detect contexts and learn habits automatically with the least possible intervention of users. In order to achieve this, machine learning (ML) techniques need to be merged at the core of reasoning models. These techniques offer powerful tools to automatically detect patterns, categorize contexts, build usage profiles, represent data with compact mathematical hypothesis and provide statistical information vital for the intelligent aspect of AmI. This dissertation ends up by opening new directions on how to model and adapt machine learning techniques to fit for AmI platforms.
Overall, this thesis provides solutions for the next leap of technology, where sensors become ubiquitous. Our solutions, implemented in an open source framework KMF, allow to create efficient and distributed, data and component models for IoT, adaptable at runtime leveraging multi-objective optimization to find good tradeoff between qualities for the current context, and machine learning techniques to derive contextual rules, profile and learn habits automatically.
