Upravljanje kinematskim strukturama inspiriranih prirodom

master's thesis

• Title (english): Control of nature-inspired kinematic structures
• Author: Marko Ivić
• Mentor: Domagoj Jakobović (mentor)
• Committee member: Domagoj Jakobović (predsjednik povjerenstva)
• Committee member: Marko Đurasević (član povjerenstva)
• Committee member: Ante Đerek (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-07-05, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Upravljanje kinematskim strukturama je zahtjevan i kompleksan problem. Uz pomoć robusnog i brzog pokretača fizike MuJoCo moguće je stvoriti okolinu za podržano učenje ponašanja kinematske strukture. Ponašanje se uči putem različitih funkcija dobrote ovisno o željenom ponašanju. Kinematskim strukturama se upravlja putem neuronske mreže koja na svoje ulaze dobiva podskup podataka iz trenutnog stanja simulacije, a na izlazu daje pomake aktuatora. Algoritmi učenja optimiraju težine neuronske mreže kako bi maksimizirale funkciju dobrote. Korišteni algoritmi učenja su: genetski algoritam s dva para operatora križanja i mutacije, algoritam roja čestica i evolucijska strategija. Svi algoritmi u nekoj mjeri pokazuju učenje inteligentnog ponašanja. Ponašanja koja algoritmi nauče znaju biti slična intuitivnim ljudskim ponašanjima.

Abstract (english)

Control of kinematic structures can prove to be a difficult and complex problem. MuJoCo is a robust and fast physics engine which enables creating the environment for reinforcement learning. Fitness function is used to learn the desired behavior of a kinematic structure. Kinematic structures are controlled using a neural network which takes a subset of a current state of the simulation and outputs actions for the actuators in the physics model. Weights in neural networks are optimized using evolutionary learning algorithms with a goal to maximize fitness function. Learning algorithms used: genetic algorithm, evolution strategy and particle swarm optimization. All algorithms show intelligent behavior to some degree. Learned behaviors can be similar to a human when completing the same task.

• Keywords: evolucijsko računarstvo; genetski algoritam; algoritam roja čestica; MuJoCo; kinematska struktura; neuronska mreža; podržano učenje; evolucijska strategija
• Keywords (english): evolutionary algorithms; particle swarm optimization; genetic algorithm; evolution strategy; MuJoCo; artificial neural network; reinforcement learning
• Language: croatian
• URN:NBN: urn:nbn:hr:168:911415
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-14 08:57:32