/library/oar/handle/123456789/63780 2025-12-28T13:04:35Z /library/oar/handle/123456789/121368 Title: MIRAI : a modifiable, interpretable, and rational AI decision support system Abstract: With the recent advancements and results obtained by Deep Learning, several corporations are eager to begin incorporating these algorithms into their workflow to benefit from these systems, especially with the emergence of Industry 4.0. However, decision makers using these systems find themselves unable to fully trust the AI from evaluation metrics alone, and require some more transparency in the rationale behind their systems. As such, research has gone in the direction of Explainable AI (XAI), where the creation of explainable algorithms or reverse-engineering of existing algorithm takes place to open up the Black Box of opaque AI algorithms. Both approaches present the outcome in a manner which is interpretable by humans. In this research project, we proposed an Explainable AI architecture for predictive analysis in industry. We intended to utilize a novel approach of combining the Rule-Based Reasoning methodology of the Differentiable Inductive Logic Programming (δILP) algorithm with an explainable Machine Learning (ML) algorithm, a Bidirectional Long-Short Term Memory (BiLSTM) Neural-Network (NN). The combination of these algorithms created a fully explainable system capable of a higher level of reasoning, or ’Deep Understanding’. In turn, this implementation of Deep Understanding allowed us to have more reliable and faithful explanations for the given application. Quantitative evaluation for this system took place by means of standard Machine Learning evaluation metrics such as F1-Scores, Precision, Recall, and Receiver Operating Characteristic (ROC) Curves. Our BiLSTM performed with an average of 85% over several metrics, and δILP performed at over 95%. We further evaluated our system by taking the derived interpretations, transforming them into English explanations via the inferences and deductions stored in a standardized Knowledge Base. We verified them with industry professionals to determine whether the deductions made sense in a practical context or not. From this, we understood that a combination of values in the acceleration and rotation in either the X and Y axis exclusively may lead to an error. Highlighting these features in an explanation and sorting them by their strength gives technicians an idea of what solution to apply when met with this explanation, saving time in deconstructing the problem when met with it, and in turn improving Overall Equipment Efficiency (OEE). In future work, we would enhance these explanations by moving our solution to prescriptive maintenance, where we also highlight possible solutions for the error indicated by MIRAI. Description: M.SC.ARTIFICIAL INTELLIGENCE 2020-01-01T00:00:00Z /library/oar/handle/123456789/85570 Title: Learning DFAs from sparse data Abstract: Regular inference is the task of identifying a Deterministic Finite-State Automaton (DFA) from a training set of positive and negative strings of a regular language. This problem is known to become increasingly more difficult as the size of the target DFA grows larger and as the training data becomes sparser. One of the most studied algorithms for this task is the Evidence-Driven State Merging (EDSM) algorithm due to Rodney Price which emerged as a winning algorithm in the international Abbadingo One competition of 1997. We focus on ‘Abbadingo-style’ problems (learning DFAs over binary alphabets with training sets of various densities), and we present the results of a comprehensive analysis of how, and more importantly when, EDSM succeeds or fails in identifying either the exact target DFA, or a low-error hypothesis with respect to a test set. We do this by creating thousands of problem instances to study their characteristics, as well as the behaviour of the state merging process. To support this analysis, we have developed an extensive software framework consisting of highly optimised, and parallelised, implementations of state merging algorithms, as well as several visual and statistical analysis tools. Motivated by the results and insights we obtained following this analysis, we propose three methods each having the aim of improving on the generalisation rate of EDSM on Abbadingo-style problems when the training data is sparse. Our first method involves the development of an ensemble of monotonic, greedy heuristics which, together, are able to outperform EDSM. This method is inspired by Wolpert and Macready’s No Free Lunch theorems which roughly state that, in general, no single heuristic will dominate all others over all problem instances. This is indeed supported by the empirical evidence we have gathered during our experimentation on various problem configurations. Associated with the ensemble of heuristics, we have also identified a method which enables us to predict, with a high degree of confidence, which of the individual heuristics in the ensemble results in a low or zero-error hypothesis. Our second approach, which we call the Delta Graph method, is based on the observation that when a greedy heuristic selects sequences of merges, the initial ones are especially critical. When a wrong choice is made, finding the target DFA becomes impossible and the likelihood of finding a low-error hypothesis will be greatly reduced. This method involves constructing and non-monotonically searching in a structure representing a highly condensed subspace of possible merges. This structure contains several short sequences of merges, where, with high experimental probability, at least one of them will consist of correct merges leading to the target. These merges establish enough constraints on a partial hypothesis that, when extended with a label-matching heuristic, will lead to the target DFA or a low-error hypothesis. Typical evolutionary approaches in DFA learning operate by attempting to evolve a target DFA either as a transition matrix or by partitioning the states of a Prefix Tree Acceptor (PTA). In our third method, we present an alternative approach which, instead, evolves short sequences of merges selected from a subset of high state-reduction merges. As in the Delta Graph method, these short sequences of merges establish enough constraints on a hypothesis, that when extended with a label-matching heuristic, will, with high experimental probability, lead to the target DFA or a low-error hypothesis. To ensure a common baseline for comparison, our methods are evaluated on target DFAs and training sets which have been constructed according to the Abbadingo One competition procedures. Our results show that each of the methods we have developed outperforms EDSM. For example, on 64-state target DFA problems and symmetrically structurally complete training sets at the sparsest density set by the Abbadingo One competition, while EDSM identifies low-error DFAs approximately 15% of the time, our ensemble, Delta Graph, and evolutionary methods do so about 26%, 43%, and 56% of the time respectively. We also obtain considerably better generalisation rates on problem instances which are highly adversarial to EDSM. Description: PH.D.ARTIFICIAL INTELLIGENCE 2020-01-01T00:00:00Z /library/oar/handle/123456789/64245 Title: Named entity recognition for Maltese : a scenario for a low resource language Abstract: Named Entity Recognition (NER) is a subtask in the NLP field whereby named entities such as Person, Organization and Location are identified and labelled in text. NER is a huge contribution to ����������� Extraction as it identifies the named entities from which the needed information can be extracted, such as their relations also known as entity linking. Generally, state-of-the art NERs are trained on large corpora with Named Entities already tagged through human annotation initiative. However, not all languages have such huge corpora available. In fact, Maltese has neither available NER annotated datasets nor previously created Maltese NER models. Hence, the aim of this study is to conduct research about previous low-resource NERs to obtain enough knowledge to solve the challenging task of creating and evaluating the first Maltese NER system. For this research, we created a small dataset of 500 sentences extracted from a publicly available Maltese corpus and manually annotated them with Person, Location, Organization and Miscellaneous entities using the BIO tagging system. In order to augment our dataset, we experiment with transfer learning by including datasets from other languages which are English, Italian, Spanish and Dutch as they can be deemed as rather similar to Maltese. Our experiments evaluate the use of two techniques which are Conditional Random Fields (CRF) and Bidirectional Long Short-Term Memory Conditional Random Fields (BiLSTM-CRF) as a deep learning approach. Our experiments also had to consider a number of scenarios since there were no specific annotation guidelines for Maltese. Initially, tags were limited to Person, Organisation and Location, with a later introduction of the Miscellaneous tag for further experimentation. The analysis of the tags had to also match what is available in the selected multilingual NER datasets to make the transfer learning streamlined to the Maltese annotations. We also experimented with the size of the multilingual corpora to analyse the impact that other languages can have on Maltese NER. This is done incrementally with the first corpus containing Maltese only and then the others contain one of the following amounts of sentences from each language: 200, 300, 400 and 500. These experiments resulted in a large number of setups, totalling to 40 distinct experiments. The best results are obtained by three equally successful systems achieved by the BiLSTM-CRF’s experiments. One of these systems is trained on Maltese and 300 extra sentences from the other languages without making use of the Miscellaneous tag. The other two systems are trained on Maltese together with 400 and 500 extra sentences from each language but excluding Dutch. Description: B.SC.ICT(HONS)ARTIFICIAL INTELLIGENCE 2020-01-01T00:00:00Z /library/oar/handle/123456789/64244 Title: GAImE : Investigating game AI for enhanced user experience Abstract: Over the last few decades, the business of digital and video games has risen exponentially. Aside from the higher fidelity of the graphics and the more involving game plots and narratives, GameAI is an area which is gaining more popularity in the game development industry. Research and development in the area is registering increased progress in a number of AI technologies. Game AI research that has been ongoing for the past two decades has also seen an increase into the investigation of how AI can be used to enhance the user’s game play experience rather than just learning how to beat human players. The concept in applying AI techniques is to make NPCs (Non-Player Characters) appear intelligent by making use of different techniques. NPCs are AI-controlled characters within the video game which can have different typologies such as companions and enemies. This thesis investigates different AI techniques for the generation of NPCs and how these may in turn affect the user’s experience by evaluating user engagement and satisfaction. Game studies indicate that making use of eye-catching graphics, great sound effects and music, are essential game elements and may improve user satisfaction. However, in this thesis, we propose an additional element that may add to the user satisfaction in game play – intelligence in NPCs. As a proof of concept, a survival game is developed using Unity 3D. Different AI techniques, such as FSM, Behaviour Based NPCs and RL Agents have been adopted and adapted to the game’s NPCs and these have been in turn tested and evaluated from the user experience perspective. Whilst for behaviour-based NPC’s the actions were predefined, RL agents learned actions depending on the in-game rewards obtained. The satisfaction of the user is calculated by using the Game User Experience Satisfaction Scale. This gives a good indication of the parts of the game which make the user feels more satisfied. The results record a total average satisfaction of 5.8/7 on the GUESS scale. The survey results also indicated that users preferred the behaviour-based approach in favour of RL. This could be due that for these particular group of users preferred to predict and anticipate the actions and reactions of the behaviour-based companion NPC since all of its actions were predefined. These preferences are subjective depending on the type of audience playing the game. Description: B.SC.ICT(HONS)ARTIFICIAL INTELLIGENCE 2020-01-01T00:00:00Z