Enhancing Tetris Gameplay with Deep Reinforcement Learning
Popek, Dawid (2024)
2024
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2024061022571
https://urn.fi/URN:NBN:fi:amk-2024061022571
Tiivistelmä
This work's principal purpose was to create an artificial intelligence solution for the well-known Tetris game, which was totally constructed in C++ since there aren't many online resources in this subject. Furthermore, C++ was utilized to develop the reinforcement learning agent, assuring compatibility, and enabling the simple integration of the LibTorch library.
Deep Q-Network, a deep reinforcement learning system initially described by DeepMind in 2015, is the method employed in this thesis. A prioritized experience replay, in which transitions with the highest temporal difference mistake are sampled more often, was utilized in replace of the original algorithm's conventional experience replay memory.
The AI agent made an effort, but its performance was not fully pleasing as it was unable to find out the optimum approach for playing the game. However, this study has built a good platform for additional investigation. By adjusting the hyperparameters and the reward function, the results could be improved.
This thesis addressed the absence of freely available solutions for Tetris reinforcement learning settings and DQN algorithms written in C++ on the internet, by developing a full base that researchers and enthusiasts may utilize in their projects.
Deep Q-Network, a deep reinforcement learning system initially described by DeepMind in 2015, is the method employed in this thesis. A prioritized experience replay, in which transitions with the highest temporal difference mistake are sampled more often, was utilized in replace of the original algorithm's conventional experience replay memory.
The AI agent made an effort, but its performance was not fully pleasing as it was unable to find out the optimum approach for playing the game. However, this study has built a good platform for additional investigation. By adjusting the hyperparameters and the reward function, the results could be improved.
This thesis addressed the absence of freely available solutions for Tetris reinforcement learning settings and DQN algorithms written in C++ on the internet, by developing a full base that researchers and enthusiasts may utilize in their projects.