Akademska digitalna zbirka SLovenije - logo
VSE knjižnice (vzajemna bibliografsko-kataložna baza podatkov COBIB.SI)
  • Deep reinforcement learning compared to human performance in playing video games [Elektronski vir]
    Markovska, Jovana, 1999- ; Šoberl, Domen
    Through deep reinforcement learning, a computer can learn to play simple video games under the same conditions as human players - by watching the pixels on the screen and issuing the button-press ... type of actions. In this paper, we investigate how quickly a computer can reach and surpass human performance in a simple two-player video game, given no background knowledge of how to play the game. We implement a Deep Q-Learning (DQN) algorithm using Python and integrate it with the Atari 2600 emulator that runs the Pong game. We train the neural network for up to 3 million training steps and evaluate its performance after every 100.000 steps. As a reference, we measure the performance of 18 human players and take their average rating as the human-level baseline. We propose an evaluation metric that considers the obtained game points and the player's endurance during the game. We find that the Deep Q-Learning algorithm can surpass beginner-level human players in playing the Pong game after about 4 hours of training.
    Vrsta gradiva - prispevek na konferenci
    Leto - 2022
    Jezik - angleški
    COBISS.SI-ID - 134386179