Research On Context-Aware Computing Technology

Ubiquitous/Pervasive Computing is believed to be the next computing paradigm, characterized by pervasiveness and transparency, to deliver personalized services; while context-aware computing is thought to be the only way leading to the wonderful vision of ubiquitous computing. Considering that research on context-aware computing is on its preliminary stage, this thesis made researches on three aspects of the infrastructures of context-aware computing and one application. Specifically,1. As the basis of context-aware computing and our subsequent studies, an unwatched automatic algorithm, based on Fuzzy Sets, is propsed to discretize continuous contexts that follows approximate Gaussian distribution. This algorithm is built up of a fact that many kinds of continuous contexts, in practice, conform to approximate Gaussian distribution, which can be fitted by a sum of Gaussian distributions. Experiments show that this algorithm was effective; hence, it will be referred in later chapters.2. A situation-aware approach is brought forward to deal with the inconstancy and uncertainty problems inevitable in context-aware computing. Formal definitions about inconstancy and uncertainty are given by this thesis, and uncertainty is explained as incompleteness, inaccuracy and inconsistency. Bayesian network, modelling causal relationships between contexts and situations, is employed to facilitate our situation-aware approach, which in turn helps shield inconstancy. Algorithms based on Expectation-Maximization is built to handle incompleteness and inaccuracy; while ontology technology is used to eliminate inconsistency. Simulations reveal that this situation-aware approach outperforms context-aware approach in user interruption ratio and work efficiency without losing accuracy.3. Context reduction and rule generation methods are exploited to dealing with the trivialness problem common in context-aware computing. Again, Bayesian network is first employed to model causal relationships between contexts and situations; then, Markov blanket is applied to this Bayesian network to find the core contexts, which is less than full contexts. This procedure is context reduction. After that, to compensate those contexts, excluded by the core contexts, in identifying situations, a "two-step" rule generation method is derived from the core contexts to identify situations. Experiments prove that those methods are correct and effective; specifically, trivialness can be eliminated exponentially, if not completely.4. A systematic approach for context-aware end-to-end QoS qualitative diagnosis and quantitative guarantee is studied as an illustration of context-aware application. Bayesian network structure learning and parameter learning are employed to exploit qualitative and quantitative causal relationships between contexts and QoS metrics, separately. QoS qualitative diagnosis is reached according to the learned structure; while QoS quantitative guarantee is realized by tuning tunable causal contexts to their quantitative values derived from learned parameters, respectively. Experiments validate that this approach is effective; furthermore, in general, this approach can be reached with a polynomial time complexity in practice. Even if in the worst scenarios, the time complexity is still hold if contexts are reduced in advance by the context reduction method mentioned above, or if the total number of contexts and QoS metrics is assured in the order of magnitudes of ten.