Deep reinforcement learning for games using adamdql (deep q-learning) = deep reinforcement learning untuk permainan dengan metode adam-dql ( deep q-learning)

Halim, Andrew Mahisa (2018) Deep reinforcement learning for games using adamdql (deep q-learning) = deep reinforcement learning untuk permainan dengan metode adam-dql ( deep q-learning). Bachelor thesis, Universitas Pelita Harapan.

[img] Text (Title)
Title.pdf
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (2MB)
[img] Text (Abstract)
Abstract.pdf
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (588kB)
[img] Text (ToC)
ToC.pdf
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (694kB)
[img] Text (Chapter 1)
Chapter 1.pdf
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (448kB)
[img] Text (Chapter 2)
Chapter 2.pdf
Restricted to Registered users only
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (391kB)
[img] Text (Chapter 3)
Chapter 3.pdf
Restricted to Registered users only
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (1MB)
[img] Text (Chapter 4)
Chapter 4.pdf
Restricted to Registered users only
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (1MB)
[img] Text (Chapter 5)
Chapter 5.pdf
Restricted to Registered users only
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (845kB)
[img] Text (Chapter 6)
Chapter 6.pdf
Restricted to Registered users only
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (352kB)
[img] Text (Bibliography)
Bibliography.pdf
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (361kB)
[img] Text (Appendices)
Appendices.pdf
Restricted to Repository staff only
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (6MB)

Abstract

The field of deep learning has gained a huge traction over the last few years. Its youngest sub-field, Deep Reinforcement learning (RL) has shown remarkable potential for artificial intelligent based opponent in games. Several research based on Deep RL begun to appear and ultimately lead to the development of Deep Q-Learning, a deep learning technique that allows an agent to learn from an image without the help of human-created model or features. Since Deep Q-Learning (DQL) is still in its early stages, it was mostly tested on simple, toy-like examples. This thesis will try to take a step further and apply deep reinforcement learning to complex games. This will be achieved by combining classical DQL with Adam optimizer, and several policy improvement techniques. This thesis also introduced partial training, a policy improvement technique for neural network that kickstarts an agent to get rewards faster in complex games. Adam-DQL agent is then tested on game environments based on real life video games. The results indicate that Adam-DQL agent is learns faster and performs significantly better compared to classical Deep Q-Learning. We further shows that combined with partial training, Adam-DQL is viable to solve even really complex games./ Bidang Ilmu pembelajaran dalam (deep learning) sedang meledak popularitasnya selama beberapa tahun terakhir. Cabang ilmunya yang termuda, deep reinforcement learning, telah menunjukan potensi yang luar biasa sebagai lawan bermain berbasis kecerdasan buatan di permainan video. Beberapa riset dalam bidang ini mulai bermunculan dan berujung pada pengembangan Deep QLearning, teknik pembelajaran dalam yang memungkinkan sebuah agen kecerdasan buatan untuk belajar hanya dari gambar dalam permainan, tanpa campur tangan manusia. Karena Deep Q-Learning (DQL) masih dalam masa pengembangan awal, teknik ini umumnya diaplikasikan pada permasalahan yang sederhana. Skripsi ini akan mencoba melangkah lebih jauh dengan mengaplikasikan Deep Q-Learning pada permainan yang kompleks. Hal ini dapat dicapai dengan mengkombinasikan Deep Q-Learning klasik dengan teknik optimisasi Adam, dan beberapa teknik perbaikan kebijakan (policy improvement techniques). Skripsi ini juga memperkenalkan partial training, teknik perbaikan kebijakan yang membantu agen permainan untuk mendapatkan lebih banyak ganjaran (reward) di awal permainan pada permasalahan yang lebih kompleks. Adam-DQL kemudian diuji pada lingkungan permainan yang dibuat berdasarkan permainan video popular di masyarakat. Hasilnya mengindikasikan bahwa Adam-DQL belajar lebih cepat serta memiliki performa bermain yang jauh lebih baik dibandingkan teknik Deep Q-Learning klasik. Kemudian, dapat ditunjukkan bahwa jika dikombinasikan dengan partial training, Adam-DQL bahkan mampu menyelesaikan permainan yang sangat sulit.

Item Type: Thesis (Bachelor)
Creators:
CreatorsNIMEmail
Halim, Andrew MahisaUNSPECIFIEDUNSPECIFIED
Contributors:
ContributionContributorsNIDN/NIDKEmail
Thesis advisorLukas, SamuelUNSPECIFIEDUNSPECIFIED
Thesis advisorMargaretha, HelenaUNSPECIFIEDUNSPECIFIED
Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Divisions: University Subject > Current > Faculty/School - UPH Karawaci > School of Information Science and Technology > Informatics
Current > Faculty/School - UPH Karawaci > School of Information Science and Technology > Informatics
Depositing User: Mr Samuel Noya
Date Deposited: 02 Oct 2019 03:51
Last Modified: 02 Oct 2019 03:51
URI: http://repository.uph.edu/id/eprint/4723

Actions (login required)

View Item View Item