A Generative Policy Gradient Approach for Learning to Play Text-Based Adventure Games
Text-based adventure games pose a very difficult task for reinforcement learning. Current approaches work well on generated games but are limited to specific (sets of) games only and require knowledge about possible actions in advance. The master’s thesis presented here describes an early first step towards a more generative and general approach of learning to play text-based games by moving from discrete actions to learned continuous action representations.
René Raab graduated from Maastricht University in 2019 receiving an M.Sc. in Artificial Intelligence. He is now a PhD candidate at Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) in Germany.