Imitating Agents in A Complex Environment by Generative Adversarial Imitation Learning
The generative adversarial imitation learning (GAIL) shows the ability to find reward functions to explain expert players' behaviors in some low-dimensional environments using hand-crafted features as inputs. In this research, we aim to extend GAIL to complex environments and using raw images a...
Saved in:
Published in | 2020 IEEE Conference on Games (CoG) pp. 702 - 705 |
---|---|
Main Authors | , , |
Format | Conference Proceeding |
Language | English |
Published |
IEEE
01.08.2020
|
Subjects | |
Online Access | Get full text |
Cover
Loading…
Summary: | The generative adversarial imitation learning (GAIL) shows the ability to find reward functions to explain expert players' behaviors in some low-dimensional environments using hand-crafted features as inputs. In this research, we aim to extend GAIL to complex environments and using raw images as inputs. We propose to (1) use convolutional neural networks to deal with image inputs, (2) adopt a structure called global-local discriminator to GAIL, and (3) represent trajectories as state-state pairs instead of state-action pairs. Our approach successfully imitates given players in Super Mario Bros. To our knowledge, the results are the first to have successful imitations in complex environments based on image inputs. |
---|---|
ISSN: | 2325-4289 |
DOI: | 10.1109/CoG47356.2020.9231805 |