?url_ver=Z39.88-2004&rft_id=10.5204%2Fthesis.eprints.230761&rft_val_fmt=info%3Aofi%2Ffmt%3Akev%3Amtx%3Adc&rft.title=Efficient+and+stable+reinforcement+learning+for+robotics&rft.creator=Dasagi%2C+Vibhavari&rft.subject=Reinforcement+Learning&rft.subject=Exploration&rft.subject=Artificial+Curiosity&rft.subject=Robot+Learning&rft.subject=Deep+Learning&rft.description=Reinforcement+Learning+(RL)+has+long+been+used+for+learning+behaviour+through+agent-collected+experience%2C+recently+boosted+by+deep+neural+networks.+However%2C+typical+deep+RL+agents+require+millions+of+training+data+samples%2C+equating+to+days+or+weeks+of+training+in+simulation%2C+and+months+to+years+in+the+real+world.+As+robot+experience+is+expensive%2C+this+magnitude+of+real+robot+training+is+not+desirable.+In+this+thesis%2C+we+address+the+issue+of+efficiency+in+RL+to+make+it+feasible+option+for+robot+learning+in+the+real+world+by+focusing+on+improvements+in+three+key+aspects%3A+data+collection%2C+data+usage+and+policy+training.&rft.publisher=Queensland+University+of+Technology&rft.date=2022&rft.type=Thesis&rft.format=application%2Fpdf&rft.relation=https%3A%2F%2Feprints.qut.edu.au%2F230761%2F1%2FVibhavari_Dasagi_Thesis.pdf&rft.rights=free_to_read&rft.rights=http%3A%2F%2Fcreativecommons.org%2Flicenses%2Fby-nc-nd%2F4.0%2F&rft.relation=doi%3A10.5204%2Fthesis.eprints.230761&rft.relation=Dasagi%2C+Vibhavari+(2022)+Efficient+and+stable+reinforcement+learning+for+robotics.+PhD+thesis%2C+Queensland+University+of+Technology.&rft.id_number=https%3A%2F%2Feprints.qut.edu.au%2F230761%2F&rft.identifier=Faculty+of+Engineering%3B+School+of+Electrical+Engineering+%26+Robotics