• Nichlas Ørts Lisby
  • Thomas Højriis Knudsen
  • Tobias Lambek Jacobsen
  • Nicolaj Casanova Abildgaard
3. semester, Software, Kandidat (Kandidatuddannelse)
This survey encompasses four reinforcement learning (RL) algorithms, namely the Deep Q-learning Network (DQN), Dueling Double Deep Q-Learning Network (D3QN), Deep Q-Learning from Demonstrations (DQfD) and Advantage Actor Critic (A2C) algorithms.
The survey is conducted with the purpose of gaining knowledge in the field of RL, and specifically to understand how different algorithms are developed and how they compare to one another.
Each algorithm is reproduced and evaluated upon.
We develop a generic framework which allow us to remain consistent and fair in the training and experimentation of these algorithms.
We demonstrate the algorithms performance across 4 different Atari 2600 games and evaluate the results in terms of performance and efficiency.
We are successful in reproducing three of the four algorithms, of which we achieve scores comparable to those originally reported.
SprogEngelsk
Udgivelsesdato11 jan. 2021
Antal sider58
ID: 400473471