apodolskiy / textworld_challenge Goto Github PK

Run several games manually just to get hang on what agent is supposed to do.
Tasks could be group by their toughness, type of actions, etc.
The more information we get better it would be.
Also if we group games in several clusters. It would be easy to create or leave some
games for validation.

Implement or obtain Oracle agent for training games.

1. Deep Reinforcement Learning in Text-based Games:
   https://github.com/nik31096/articles/blob/master/Language_understanding_for_text_based_games_using_DRL.pdf
   https://github.com/nik31096/articles/blob/master/Deep_RL_with_Natural_language_action_space.pdf
2. Affordances topic:
   https://github.com/nik31096/articles/blob/master/nlp4rl.pdf
   http://vision.stanford.edu/pdf/zhu14.pdf

One epoch on the lstm-dqn agent takes too long time for the whole train games dataset. Code should be proffiled to enlighten most inefficient parts.