Please use this identifier to cite or link to this item:
Title: Aprenentatge profund per reforç aplicat al control automàtic de la locomoció de robots bípedes simplificats en entorns simulats
Author: Castaño Ribes, Rafael Jesús
Director: Kanaan Izquierdo, Samir
Tutor: Ventura Royo, Carles  
Keywords: Deep Reinforcement Learning
Bipedal Locomotion
Simulated Robotics
Issue Date: 5-Jan-2021
Publisher: Universitat Oberta de Catalunya (UOC)
Abstract: CONTEXT: The subject of this Master's Thesis is the autonomous control of biped locomotion by means of artificial intelligence. PURPOSE: To know the state of the art of this field and to implement a modern solution to this problem in a simplified software-simulated environment. METHODOLOGY: 1) The problem is described in detail and related to the (Deep) Reinforcement Learning (DRL) field. 2) Theoretical foundations of DRL and its main methods applicable are analyzed. 3) The OpenAI Gym platform and its environments are analyzed; They are accepted as the platform on which to develop the product. 4) Different libraries available for DRL are analyzed and one is chosen (the TF-Agents library). 5) The algorithm to be implemented is chosen (the NAF algorithm), and the product to be developed is designed. The design includes a set of tools needed for its operation. A wrapper is also developed to compact series of observations of the environment, based on Mnih et al. (2015) in their DQN vs Atari experiment. 6) The designed product is implemented in Python. RESULTS: 1) The implemented wrapper has a positive effect on the learning of the agents. 2) The developed agent works properly and is able to solve the problem when combined with the wrapper. CONCLUSIONS: * DRL is a complex discipline, especially when the action space of the problem is continuous. * The problem can be approximatedly solved using DRL. * The TF-Agents library, although under development, has been very useful in deepening in the knowledge of DRL and its components.
Language: Catalan
Appears in Collections:Bachelor thesis, research projects, etc.

Files in This Item:
File Description SizeFormat 
rcastariTFM0121.pdf2.94 MBAdobe PDFView/Open
producte.zip137.07 kBUnknownView/Open
rcastariTFM0121Presentacio.pdf1.26 MBAdobe PDFView/Open
videos.zip10.56 MBUnknownView/Open

This item is licensed under a Creative Commons License Creative Commons