Banca de QUALIFICAÇÃO: ANDRÉ LUIZ AGUIAR SOUSA

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : ANDRÉ LUIZ AGUIAR SOUSA
DATE: 06/12/2019
TIME: 08:30
LOCAL: Sala DCA 01
TITLE:

A NEURO-FUZZY APPROACH FOR DEEP REINFORCEMENT LEARNING SYSTEMS

KEY WORDS:

Reinforcement Learning, Q-learning, Deep Reinforcement Learning,  Neuro-fuzzy systems

PAGES: 51
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUBÁREA: Eletrônica Industrial, Sistemas e Controles Eletrônicos
SPECIALTY: Controle de Processos Eletrônicos, Retroalimentação
SUMMARY:

Reinforcement learning algorithms provide a systematic way of coding behaviors deeply rooted in animal behavior psychology and neuroscience, of how an agent can improve the choice of his actions in a given environment. However, traditional algorithms run into real-world complexity, agents must be able to recognize complex environments from large sensory inputs and use them to generalize past experience to new situations. Although reinforcement learning agents have achieved some success in a variety of domains, their applicability has previously been limited to domains with low-dimensional state spaces. With recent advances in deep neural network training, especially deep Q-network, current algorithms are able to learn efficient policies even in domains where sensory data is large. However, the function of a neural network is similar to a black box, when the network fails some control tasks it is very difficult to manually adjust the network. On the other hand, fuzzy systems have the characteristic of interoperability and efficient processing of inaccurate and noisy data. Thus, this work aims to explore the use of neuro-fuzzy systems in conjunction with the recent deep Q-network training techniques to obtain agents that behave ever closer to a human level.

BANKING MEMBERS:
Presidente - 1451883 - FABIO MENEGHETTI UGULINO DE ARAUJO
Interno - 1328152 - CARLOS EDUARDO TRABUCO DOREA
Externo à Instituição - JOSE MEDEIROS DE ARAUJO JUNIOR - UFPI