This project implements an artificial agent for cometing at the TextWorld challenge (aka.ms/textworld)
The file mctsagent.py defines a Monte Carlo Tree Search agent that solves TextWorld games. It uses PUCT search, which in turn leverages a neural network to evaluate unseen game states.
The file attention.py defines the neural network that takes as input an array of observations (or memory) and learns to predict the value of the game state given those observations, a policy for the next decision, and a language model to generate commands.