@inproceedings{oai:ipsj.ixsq.nii.ac.jp:00199987,
 author = {Tianshuai, Yu and Yoshimasa, Tsuruoka and Tianshuai, Yu and Yoshimasa, Tsuruoka},
 book = {ゲームプログラミングワークショップ2019論文集},
 month = {Nov},
 note = {Significant progress has been made in the field of Reinforcement Learning (RL) in recent years. Using artificial neural networks, researchers are able to train agents that can play video games as well as or even better than human experts. However, it is common that the same environments are used in both training phases and testing phases, which results in agents’ failure to generalize to other environments. In this work, we propose a method in which environment models and generative models are used to generate virtual game levels so as to improve the generalization performance of RL agents. We conducted experiments using a fully-observable deterministic discrete maze game in order to test the proposed method. However, the proposed method failed to converge during training because our environmnet model was not able to predict the future of unseen levels accurately., Significant progress has been made in the field of Reinforcement Learning (RL) in recent years. Using artificial neural networks, researchers are able to train agents that can play video games as well as or even better than human experts. However, it is common that the same environments are used in both training phases and testing phases, which results in agents’ failure to generalize to other environments. In this work, we propose a method in which environment models and generative models are used to generate virtual game levels so as to improve the generalization performance of RL agents. We conducted experiments using a fully-observable deterministic discrete maze game in order to test the proposed method. However, the proposed method failed to converge during training because our environmnet model was not able to predict the future of unseen levels accurately.},
 pages = {150--154},
 publisher = {情報処理学会},
 title = {An Attempt to Improve Generalization Performance in Reinforcement Learning with Deterministic World Models and WGANs},
 volume = {2019},
 year = {2019}
}