Μεταπτυχιακές Διατριβές

Μόνιμο URI για αυτήν τη συλλογήhttps://dspace.library.tuc.gr/handle/123456789/82

Νέα

134

Περιηγούμαι

Πρόσφατες Υποβολές

Τώρα δείχνει 1 - 2 από 2
  • Δημοσίευση
    Grammatical inference for event recognition
    (Πολυτεχνείο Κρήτης, 2014) Kofinas Nikolaos; Κοφινας Νικολαος; Lagoudakis Michael; Λαγουδακης Μιχαηλ; Bletsas Aggelos; Μπλετσας Αγγελος; Garofalakis Minos; Γαροφαλακης Μινως
    As robot technology finds applications in the real world (search and rescue, daily household tasks, etc.), huge amounts of data are generated during autonomous robot missions. In such applications, it is often desirable to recognize high-level events that may have occurred during a mission either online or offline. Event Recognition in robot missions currently relies on human expertise and time-consuming data annotation. A modern method to recognize events is to employ Probabilistic Context-Free Grammars (PCFGs), which are formal models that can capture complex patterns in discrete sequences and can be used to parse incoming sensor data streams in order to detect patterns that may signal the occurrence of some event of interest. Recent experimentation with such methods on data from Autonomous Underwater Vehicle (AUV) missions indicated that interesting events can be recognized by parsing sequences of sensor data using an intuitive hand-written PCFG. This thesis introduces a generic procedure which can be used to automatically construct PCFGs which encode sensor data sequences that typically appear during normal robot operation using recorded logs from past missions. The resulting PCFGs can be used to recognize abnormal events in new missions evidenced by sensor data sequences which cannot be interpreted as normal. The proposed procedure consists of two parts: (a) the transformation of sensor streams into discrete sequences either to form a training corpus offline or to generate input for online parsing and (b) a Grammatical Inference algorithm in order to learn a compact PCFG consistent with a given training corpus. The learning part relies on a local search method over the space of possible grammars using chunk and merge operations. The search method aims to find a compact grammar that also maximizes its posterior probability, in a Bayesian sense, with respect to a given training corpus. The proposed procedure is evaluated on a variety of domains ranging from data-sets generated by typical context-free grammars to data-sets generated from real robot missions (NAO robot walk and AUV navigation). The results indicate that our approach is capable of producing reliable PCFG-based event recognizers, which may yield some false positive signals, but in general succeed in capturing abnormalities.
  • Δημοσίευση
    A bayesian personalized recommendation system
    (Πολυτεχνείο Κρήτης, 2014) Babas Konstantinos; Μπαμπας Κωνσταντινος; Chalkiadakis Georgios; Χαλκιαδακης Γεωργιος; Lagoudakis Michael; Λαγουδακης Μιχαηλ; Deligiannakis Antonios; Δεληγιαννακης Αντωνιος
    In this research, we propose a novel Bayesian approach for personalized recommendations. We succeed in providing recommendations that are entirely personalized, based on a user’s past item “consumptions”, building a representative user model which reflects agent’s corresponding beliefs. Having a set of items, our agent has to select the one which better matches her beliefs about a specific user, in order to recommend it and receive the corresponding reward. In our approach, we model both user preferences and items under recommendation as multivariate Gaussian distributions; and make use of Normal-Inverse Wishart priors to model the recommendation agent beliefs about user types. We interpret user ratings in an innovative way, using them to guide a Bayesian updating process that helps us both capture a user’s current mood, and maintain her overall user type. We produced several variants of our approach, and applied them in the movie recommendations domain, evaluating them on data from the MovieLens dataset. We developed a generic & domain independent system, able to face the scalability challenge and able to capture user preferences (long-term and short-term). Moreover, we dealt with the exploration vs exploitation dilemma in this domain, via the application of various exploration algorithms (e.g., VPI exploration). Ours is a completely personalized approach, which exploits Bayesian Reinforcement Learning in order to recommend an item or a top-N group of items, without the need of ratings prediction. We do not employ a Collaborative Filtering or Content-based or Preference Elicitation technique, but we are still able to provide successful recommendations. Furthermore, we tackle the famous “cold-start” problem via the use of Bayesian and VPI explorations. Our algorithms are shown to be competitive against a state-of-the-art method, which nevertheless requires a minimum set of ratings from various users to provide recommendations --- unlike our entirely personalized approach.