Do-Hyeong Kim, Imitation Leanring for Different Player Style using Generative Adversarial Networks , Master's Thesis, Department of Industrial and Systems Engineering, KAIST, 2018

Abstract

Problems that can be easily found in real life examples can usually be modeled as sequential decision problems. However they have the disadvantage that reward design is difficult when trying to solve problems using reinforcement learning methodology. In the case of Imitation learning, sequential decision problems are solved by imitating optimal behaviors using limited optimal action rather than reward. In this study, we modified the structure of VAEGAN, which is a recently developed deep generative model, to obtain the information of state efficiently when a limited amount of states and optimal action as learning data. and suggests a model that can reproduce the information as an optimal behavior. In addition, we proposed a method to conditionally learn the information about the object style and to create an action for each purpose under the same state. 

@masterthesis{Kim:2018,

author = {Do-Hyeong Kim},

advisor ={Il-Chul Moon},

title = {Imitation Learning for Different Player Style using Generative Adversarial Networks},

school = {KAIST},

year = {2018}

}