| Title |
Learning to play games with PlaNet / |
| Authors |
Valatka, Lukas |
| DOI |
10.15388/LMITT.2019 |
| ISBN |
9786090701621 |
| Full Text |
|
| Is Part of |
Lietuvos magistrantų informatikos ir IT tyrimai : konferencijos darbai, 2019 m. gegužės 14 d... Vilnius : Vilniaus universiteto leidykla, 2019. p. 70-76.. ISBN 9786090701621 |
| Keywords [eng] |
PlaNet ; Model-based reinforcement learning ; Latent space planning ; Atari gym suite ; VizDoom |
| Abstract [eng] |
An evaluation of a recent state of the art model-based reinforcement learning PlaNet in a gaming environment is presented. Author analyzes PlaNet capabilities to solve several problems in Atari and VizDoom domains. Author identifies that PlaNet’s observation and reward encoders have trouble capturing small details in Atari games (Pong, Breakout), often critical to the agent’s performance playing games. Hyperparameter tuning strategy is suggested. Author confirms latent overshooting is crucial for VizDoom Take Cover scenario, implying it is necessary for similar environments. |
| Published |
Vilnius : Vilniaus universiteto leidykla, 2019 |
| Type |
Conference paper |
| Language |
English |
| Publication date |
2019 |
