Call for Participation
If you're interested in research at the intersection of NLP and RL, join us for "First Textworld Problems: A Reinforcement and Language Learning Challenge," presented by Microsoft Research. This competition challenges researchers to devise and train an AI agent that solves simplified text-based games. It runs until June 1st, 2019.
Text-based games are interactive simulations that use natural language to describe the state of the world, to accept actions from players (AI or human), and to report subsequent changes in the environment. Through sequential decision making, players work toward goals which may or may not be specified explicitly. Text-based games are fertile ground for language-focused machine learning research because, in addition to language understanding, successful play requires skills like long-term memory and planning, exploration (trial and error), and often some common sense. Text-based games are typically tackled via Reinforcement Learning (RL).
The FTP competition evaluates agent performance across a set of related but distinct text-based games. All the games share a similar structure and theme (cooking a meal in a modern house), with similar world descriptions, similar text commands, and similar interactive entities. On the other hand, the objective, world layout, and existence/location of entities vary randomly (within constraints) across games.
We've divided the game set into training, validation, and test splits. The training set is available to download now, while we're withholding the test set for final evaluation. Participants can submit trained agents at any time to our servers to be tested on the validation set. This is a relatively new setup for RL research -- oftentimes, agents are trained and tested on the same game -- which we hope encourages the development of algorithms that generalize.
On the competition's CodaLab site (aka.ms/textworld-challenge) you'll find detailed instructions and evaluation criteria, plus a Starter Kit with two sample agents coded in PyTorch. It should be easy to get started! Join in now and devise a text-based agent for a chance to win $2000 USD and more in prizes!