ISEP - DM – Engenharia de Inteligência Artificial
Permanent URI for this collection
Browse
Browsing ISEP - DM – Engenharia de Inteligência Artificial by advisor "Faria, Luiz Felipe Rocha de"
Now showing 1 - 10 of 11
Results Per Page
Sort Options
- Ajuste dinâmico de dificuldade em videojogos usando aprendizagem automáticaPublication . Felício, Jorge Emanuel Coelho Mendonça de Anciães; Faria, Luiz Felipe Rocha deIn the constantly evolving field of video games, traditional difficulty settings fail to accommodate the wide range of skill levels among players. The resulting mismatch between the player’s skill and the game’s challenge can make the game boring for skilled players or frustrating for less experienced ones, negatively affecting player engagement. Dynamic Difficulty Adjustment (DDA) seeks to resolve this issue by adapting the game’s difficulty in real time in response to the player’s performance. While advancements in artificial intelligence (AI), particularly machine learning (ML), have enabled more adaptive DDA systems, the full potential of certain advanced techniques or tools has yet to be explored. This thesis thus explores possible innovations in the integration of AI in DDA systems for video games. The research begins by reviewing the techniques used for DDA, focusing on methodologies such as player modeling, rule-based systems, and ML. Based on this research, potential areas for innovation were identified and the application of Deep Reinforcement Learning (DRL) in the Unity game development platform through the usage of the MLAgents toolkit was chosen as a promising approach for this research. Using this methodology, this research aims to implement a DDA system that adjusts a game’s difficulty based on the player’s skills, enhancing their engagement and maintaining a consistent challenge. This project has several critical phases of development, including the creation of a game prototype, data collection for model training, development and integration of the DDA system into the game prototype, and conducting an experiment comparing the prototype with DDA integrated with a version of the prototype that used traditional static difficulty scaling. The experiment conducted was done with 20 participants of varying skill levels and used a combination of collected gameplay metrics and a modified Game Experience Questionnaire (GEQ) survey to evaluate the DDA system’s effectiveness. The results showed that the DDA system demonstrated a statistically significant increase in the player engagement component and appropriately adjusted the difficulty to be harder for participants of higher skill. However, the system sometimes exhibited some issues with drastic adjustments in difficulty between levels, which led to a slightly lower Post-Game positive experience score compared to the static difficulty scaling system. Despite these fluctuations, the proposed system demonstrates the potential of the ML-Agents toolkit in implementing DDA with DRL in games made on the Unity platform. By identifying underexplored areas in the current literature and applying advanced techniques like DRL, this thesis aims to contribute to both academic research and game development regarding the approach to DDA in video games.
- Development and evaluation of a complex-valued neural network library: the Renplex open-source projectPublication . Alves, Pedro Manuel Ferreira; Faria, Luiz Felipe Rocha deComplex-Valued Neural Networks (CVNN) have shown to be a promising type of Artificial Neural Networks (ANN) when compared to its real-valued counter-parts. However, it has been a research field where authors autonomously developed and tested CVNN with no common tools or library to module them. This Master Thesis presents a library called Renplex capable of modulating CVNN as an open-source project for research and even for small scale applications. Although not suitable for beginners in the field of ANN or programming, the library provides a low-level interactive with Machine Learning (ML) pipeline, in order to accurately control CVNN evaluation. To test the library’s core functionalities, architectures such as Complex-Valued Multi-Layer Perceptron, Auto-encoder and Convolutional Neural Network were trained. These achieved test results that outperformed their real-valued counterparts for the MNIST dataset and a synthetically generated dataset for signal reconstruction. Such improvement on performance, has been previously stated throughout literature. It consisted in greater test accuracy (or lower loss values), more stability in training, faster convergence in terms of epochs needed, greater capability of generalization, and subsequently less prone to over-fitting. This work will introduce a new tool for exploring CVNN, capable of scaling and potentially uncovering many of their hidden potentials for ML-related tasks.
- Geração de relatórios a partir da análise de avaliações de unidades hoteleirasPublication . Pereira, Isadora Manuel Almeida; Faria, Luiz Felipe Rocha deNowadays, customer reviews play a vital role in determining the success of businesses, particularly in the hospitality industry, where online feedback is both abundant and influential. However, the high volume of reviews presents a challenge for hotel owners and managers, who need efficient ways to extract useful insights. This dissertation addresses this issue by developing the FeedbackFunnel, a Natural Language Processing (NLP) model capable of analysing and summarizing customer reviews to provide concise and meaningful information. The model integrates three components: sentiment analysis, feature synthesis, and multidocument summarization. Each component was rigorously tested and improved individually to enhance performance. The Sentiment analysis was conducted using logistic regression combined with a TFIDF unigram model, chosen for its effectiveness in accurately classifying sentiments. The Feature synthesis for sentence creation component synthesized key features from sentiment analysis into sentences, summarizing the most notable positive and negative aspects of the reviews. For the summarization component, the pre-trained “sshleifer/distilbart-cnn-6-6” model was used to generate concise summaries from multiple reviews. To validate the performance of the models, traditional metrics such as accuracy were used for sentiment analysis, while more advanced measures like embedding-based similarity scores and perplexity were employed to assess the quality and coherence of the generated summaries. The developed model produced promising results by effectively capturing both positive and negative aspects mentioned in the review, even when the general sentiment leaned in one direction. However, there are still areas that can be improved. Enhancing the sentence creation component by using a pre-trained model to generate sentences could improve the coherence and richness of the generated content, moving beyond the current rigid and simplistic structure. Additionally, fine-tuning the summarization component on a domain-specific dataset could significantly improve the model’s performance.
- Machine Learning applied to forecast the outcome of professional soccer gamesPublication . Santos, João Moreira dos; Faria, Luiz Felipe Rocha deCom o aumento do poder computacional e a ênfase na Inteligência Artificial (IA) a intensificar se em diversos setores, os modelos de Machine Learning (ML) tornaram-se cada vez mais sofisticados. Considerando a notável progressão neste domínio, torna-se imperativo questionar: Num cenário hipotético em que um modelo de ML possui acesso abrangente a todas as variáveis que possam influenciar um ambiente complexo, seria o modelo de ML capaz de prever o futuro com precisão? Num cenário em que é possível criar um sistema, enriquecido com conhecimento completo de eventos passados e capacidade computacional para discernir correlações e comportamentos subjacentes, será possível prever eventos futuros com precisão? Caso seja possível, como devemos conceptualizar sorte, aleatoriedade e, em última instância, o livre arbítrio? Numa busca para investigar estas questões, esta dissertação centra-se na análise do futebol, visto ser um ambiente complexo famoso pela sua imprevisibilidade. O futebol surge como um assunto particularmente aliciante devido às suas regras estabelecidas e natureza relativamente fechada, onde a maioria das informações é conhecida antes do início dos jogos. No futebol, apesar da presença de um vasto número de variáveis exógenas, a maioria é quantificável. Dado que o futebol detém a distinção de ser o desporto mais assistido a nível global, diversas empresas capturam e disponibilizam estes dados. Ao longo desta dissertação, foi realizada uma extensa feature engineering, juntamente com uma análise detalhada do impacto de cada feature nos modelos respetivos. Foram empregues diversas metodologias de previsão, desde a Regressão Logística, previsão de séries temporais usando Autoregressive Integrated Moving Average (ARIMA) e a aplicação de Random Forests. Embora os modelos desenvolvidos nesta dissertação não tenham demonstrado conclusivamente a natureza determinística do futebol, presumivelmente devido à ausência de um conjunto de dados holístico, estes modelos superaram as previsões das casas de apostas com um rendimento de 18% para os jogos de 2021/2022 e um rendimento de 24% quando excluindo jogos com maior incerteza. Enquanto os resultados obtidos nesta dissertação não provam conclusivamente a natureza determinística do futebol, superar as casas de apostas com um rendimento satisfatório é um fator encorajador que incentiva uma melhoria futura na recolha e agregação de mais dados para possibilitar previsões ainda mais precisas.
- Near Real Time Data Aggregation for NLPPublication . Ferreira, Tiago Miguel da Costa; Faria, Luiz Felipe Rocha deCom o aumento do uso das redes sociais, o número de opções de rede para usar e a variedade de funcionalidades que elas permitem leva à necessidade de os gestores desportivos prestarem uma atenção especial a estes meios. É seguindo este pensamento que surge o Projeto PLAYOFF e consequentemente esta tese. Foi feito um levantamento da literatura existente de soluções que combinam Apache Kafka com modelos de machine learning e foi possível verificar que, apesar de soluções diferentes, já existem referencias nesses domínios. É apresentada uma comparação entre Apache Kafka e RabbitMQ e as razões da escolha ter recaído para o Kafka. É apresentada de forma geral uma arquitetura de um projeto Kafka e, posteriormente, as diferentes abordagens pensadas e desenvolvidas no âmbito da dissertação, assim como o formato das mensagens trocadas usando este sistema. Uma serie de testes e seus resultados são descritos, de modo a comprovar a sua escolha e utilização. Nestes testes diferentes abordagem de execução paralela (threads e processos) são apresentadas, assim como a forma de obter dados das APIs das redes sociais também possui diferentes abordagens. As alterações que foram realizadas aos modelos originais são descritas e explicadas as razões para essas mudanças e de que forma se enquadram na ferramenta desenvolvida. Foi realizado um teste global e final, designado por “Teste Piloto”, onde em ambiente real, com um evento real foram testados todos os componentes deste projeto, incluindo os sistemas externos desenvolvidos pela MOG Technologies e os componentes desenvolvidos no âmbito desta dissertação. Por fim, é possível comprovar as soluções apresentadas e opções finais escolhidas para o projeto, através dos resultados obtidos nos diferentes testes. É ainda proposto trabalho futuro de continuação do desenvolvido.
- Near Real-Time Sentiment and Topic Analysis of Sport EventsPublication . Albergaria, Miguel Ferraz Barbosa Soares de; Faria, Luiz Felipe Rocha deSport events’ media consumption patterns have started transitioning to a multi-screen paradigm, where, through multitasking, viewers are able to search for additional information about the event they are watching live, as well as contribute with their perspective of the event to other viewers. The audiovisual and multimedia industries, however, are failing to capitalize on this by not providing the sports’ teams and those in charge of the audiovisual production with insights on the final consumers perspective of sport events. As a result of this opportunity, this document focuses on presenting the development of a near real-time sentiment analysis tool and a near real-time topic analysis tool for the analysis of sports events’ related social media content that was published during the transmission of the respective events, thus enabling, in near real-time, the understanding of the sentiment of the viewers and the topics being discussed through each event.
- NERdy: enhancing information discovery through named entity recognitionPublication . Magalhães, João Vilas Boas da Silva; Faria, Luiz Felipe Rocha deEducation is essential for individual and societal progress, playing a pivotal role in economic development, creativity, and social mobility. However, a significant challenge remains in ensuring equitable access to quality education, particularly in diverse classrooms where personalized learning is increasingly critical. Research highlights the benefits of tailored learning approaches, but current educational tools often lack the ability to organize raw information into structured formats suitable for individualized learning. This gap underscores the need for advancements in Natural Language Processing (NLP) to enhance educational tools. In response to this challenge, this project focuses on developing a Named Entity Recognition (NER) model to improve the organization and extraction of information from raw text. NER, a key task in NLP, identifies and classifies entities such as people, organizations, and locations, providing the groundwork for future tools designed to structure educational content. The primary objective of this study is to construct an entity extraction tool, with the ultimate goal of enhancing personalized learning by facilitating the automatic organization of educational materials. To achieve this, a model combining a pretrained BERT encoder, a BiLSTM layer, and a Conditional Random Field (CRF) correction layer was developed. The model was trained on curated datasets to ensure both performance and fairness. Through extensive testing and finetuning, the model demonstrated strong results, achieving an F1 score of 87.22%, comparing favorably to state-of-the-art models. Key techniques such as class balancing, weight decay, and dropout were used to prevent overfitting, while validation and training losses were monitored to assess the model’s performance. The findings of this project not only confirm the effectiveness of the developed NER model but also highlight its potential in addressing educational challenges. The model shows promise for future expansion, including the development of relation extraction techniques and knowledge graph generation to further enhance learning tools. Ethical considerations, including data privacy, fairness, and transparency, were prioritized throughout the project. Future work will focus on refining the model and expanding its capabilities to better serve the educational sector, contributing to the broader goal of improving access to quality, personalized education.
- NLP based recommendation systems for movies platformsPublication . Silva, João dos Santos Tavares da; Faria, Luiz Felipe Rocha deNowadays, the usage of web platforms is increasingly growing. In this light, it becomes crucial, to improve the efficiency of recommendation systems in filtering and personalizing content. This thesis explores the integration of Natural Language Processing (NLP) techniques into recommendation systems with the aim of enhancing the accuracy and relevance of the suggestions provided to users. Focusing on movie recommendations, this research investigates how NLP techniques can be used to analyse user-generated content such as reviews and comments, in order to provide better compatibility between the generated recommendations and the user preferences. The aim of this thesis is to propose a model that combines sentiment analysis and topic modeling to more effectively understand and predict user preferences. The model will utilise NLP and machine learning techniques to process and analyse large datasets of user interactions and movie metadata. A comparative analysis will be presented on the use of NLP models in recommendation systems as opposed to traditional recommendation systems, highlighting improvements in accuracy and user satisfaction. The results obtained using only NLP techniques in the proposed movie recommendation model achieved an RMSE value of 0.409 and a MAE of 0.339. This result is a combination of a Logisitic Regression model with 90% accuracy and a BERTopic model with a topic coherence of 0.53. For a recommendation system based only on sentiment analysis and topic modeling of users reviews, these are satisfactory results that allow for making appropriate recommendations.
- nodeML - Towards reproducible ML in federated environmentsPublication . Silva, Edgar Simão da Mota e; Faria, Luiz Felipe Rocha deAdvances and increasing interest in AI (Artificial Intelligence) in the field of health have created novel issues, namely explainability and reproducibility of ML (Machine Learning) models. In addition, while the training of ML models traditionally favors a centralized approach, scalability and privacy issues seem to lead towards a distributed one. The latter poses challenges to ML algorithms and the efficacy of learning itself. Reproducing ML models poses several challenges arising from the intrinsic variability of the models themselves and the environment where they are trained. This problem is aggravated by their lack of standardization and common terminology. The main goal of this work is to conceptualize and prototype a framework to train, evaluate and describe ML models, in a decentralized way, over immunogenetics datasets. This framework will promote model reproducibility and comparability, as well as its adaptability. This work will start by implementing a federated/decentralized training framework over existing ML pipelines. Then, it will be possible to list and select potential dataset sources, aiming to provide an easy path to model adaptation and optimization.
- Respostas automáticas na pesquisa de conteúdos educativosPublication . Ferreira, Abel de Jesus; Faria, Luiz Felipe Rocha dePerante o constante crescimento do volume de informação disponível na internet, os motores de pesquisa desempenham um papel fundamental na forma como facilmente podemos encontrar a informação que procuramos. A utilização da Inteligência Artificial no processamento de língua natural aplicada aos motores de pesquisa de conteúdos, tem permitido que estes correspondam de forma mais inteligente, não só na vertente mais habitual para encontrar os documentos mais relevantes, mas também no entendimento das intenções do utilizador, apresentando respostas diretas às questões colocadas. Nos sistemas de eLearning, o crescimento da informação não foge à regra, acompanhando a tendência de transição digital, quer pela criação de novos conteúdos nos mais variados formatos, quer pela transformação digital dos manuais escolares. Este projeto, EVGuru, pretende explorar as técnicas mais adequadas de Inteligência Artificial aplicada ao processamento de língua natural, que permitam a criação de uma experiência de pesquisa inteligente de conteúdos educativos. O projeto é desenvolvido considerando uma eventual integração na plataforma de eLearning Escola Virtual da Porto Editora. Para tal, esta investigação é acompanhada do desenho da arquitetura dos componentes necessários à implementação de um protótipo, capaz de encontrar uma resposta direta para a questão colocada pelo utilizador, através da pesquisa nos textos de um manual escolar. Durante a realização do projeto, vários modelos foram testados em diferentes tarefas do processamento de língua natural, como classificação de textos, questões e respostas e geração de questões.