Análise de um modelo para a tomada de decisões pedagógicas em um ambiente voltado ao aprendizado de algoritmos

• Ano de defesa: 2021
• Autor(a) principal: MARANHÃO, D’jefferson Smith Santos
• Orientador(a): SOARES NETO, Carlos de Salles
• Banca de defesa: SOARES NETO, Carlos de Salles , OLIVEIRA, Alexandre César Muniz de , BARRÉRE, Eduardo
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO EM CIÊNCIA DA COMPUTAÇÃO/CCET
• Departamento: DEPARTAMENTO DE INFORMÁTICA/CCET
• País: Brasil
• Palavras-chave em Português: Tutoria inteligente; Tomada de decisões; Processos de decisão de darkov; Algoritmos de planejamento
• Palavras-chave em Inglês: Intelligent tutoring; Decision-making; Markov decision processes; Planning algorithms
• Área do conhecimento CNPq: Ciência da Computação
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/3535

Typically, algorithm teachers are responsible for presenting real-world computational problems to their students, encouraging the resolution of similar problems. This, of course, requires precious time from teachers, as it requires the selection and ordering of problems. However, the effort required by this task of organizing problems often ends up making it impractical. In this scenario, this planning stage is seen as a sequential decision-making process, in which the teacher is an agent responsible for monitoring the solutions submitted by students and presenting new problems. In particular, we have that teachers can only make assumptions about the cognitive gain provided by solving a certain problem, but they cannot measure it accurately. As a result, the present work aims to evaluate the use of POMDP models for the sequencing of computational problems, providing a more agile, dynamic and individualized form of teaching. As a way of evaluating the proposed model, randomly generated sequencing policies are compared against policies produced by specific algorithms, such as QMDP, FIB, SARSOP, POMCP and POMCPOW. The results show that the specific planning algorithms converge to very close values in terms of average discounted reward. They also indicate that a random policy always converges to a value lower than that of specific algorithms with respect to the average discounted reward. Regarding the execution time, the FIB algorithm presents the best results. The main contributions of this work are the development of a model capable of representing the pedagogical decision-making process that can be used in Intelligent Tutoring Systems (STI) aimed at learning programming; as well as the demonstration, through simulation, that planning algorithms can be applied satisfactorily to the sequencing of computational problems, surpassing the results of random policies.