The article discusses a study aimed at improving the generalization of imitation learning agents in game AI through the use of data augmentation. The authors note that while imitation learning is effective for training game-playing agents, it struggles with generalization – the ability to perform well in unseen scenarios. They propose a solution to this issue by augmenting the training data to better represent the real state action distribution. The study evaluates methods for combining and applying data augmentations to observations to improve generalization. The authors conclude that data augmentation is a promising framework for improving generalization in imitation learning agents.
Publication date: 25 Sep 2023
Project Page: Unknown
Paper: https://arxiv.org/pdf/2309.12815
