Sustentación virtual tesis doctoral Edwin Duban Torres García
El departamento de Ingeniería Eléctrica y Electrónica de la Universidad de los Andes tiene el gusto de invitarle a la sustentación de tesis doctoral de Edwin Duban Torres García que se llevará a cabo el próximo jueves 17 de diciembre a las 2 p.m. a través de zoom. A la hora indicada ingrese en el siguiente enlace
Título de la tesis: Human-like recall/association for transfer learning in Reinforcement Learning
Asesor: Fernando Lozano
Descripción:
Knowledge transfer is a feature present in the learning process of multiple animal species that machine learning algorithms are capable of imitating. Although different transfer techniques have been developed in reinforcement learning, few techniques have focused on the reproduction of the memory units involved in knowledge transfer performed by mammals, particularly the episodic memory of humans.
This dissertation presents a method that facilitates the transfer by means of a memory unit when an agent is learning to solve an unknown task that is more difficult than previously learned tasks. This dissertation focuses on developing a methodology that integrates characteristics of human learning, in terms of the structures or systems involved, in the context of reinforcement learning to make better use of the information derived from the interaction with environments in past tasks.
We show that the memory unit can be learned autonomously over a space in which situations are related through sequences of states that represent the experience of the agent. This space gives the agent a broader context for efficiently relating past experiences and determining the moments when the transfer should occur while learning a new task. We show that only previously acquired knowledge is necessary for the construction of the memory unit, which differentiates our method from those based on similarity measures, which require information on the target task to which the transfer occurs. We validate our method empirically by evaluating it in different environments where it is possible to create related tasks with increasing difficulty levels. We create different transfer scenarios in each environment and demonstrate the impact of the memory unit on the learning rate of the agent for learning new tasks.