% \section{Language based Reinforcement Learning} % Natural language has been considered recently to improve an agent’s performance and the first known instance of this was in 2016 by \cite{he:NL-space} where an agent was applied to text-based games. Following this, \cite{kaplan:NL-atari} introduced a basic approach that associates the reward function to given commands to be called for later tasks to beat previous performance in Atari games. % \cite{luketina:NL-RL-survey} provide a detailed summary of the research attempting to link NL and RL so far. It is clear that the grounding (``i.e. learning the correspondence between language and environment features") is still a significant research challenge. %\begin{itemize} % \item ``Currently, we rely on weak supervision from humans to define what skills to be learned in each training stage. In the future, we plan to automatically discover the optimal training procedures to increase the task set." \cite{shu:interp-RL}. % \item The authors take advantage of a training dataset with two human controllers; one as instructor and one as player of the game. By nature of this design, the completion of goals can be connected to the final state the human player reaches for the given instruction, thus grounding by demonstration \cite{hu:nl-instr}. %\end{itemize}