Microsoft Announces TextWorld For Training and Evaluation of RL Agents on text-based Games

Text based games have come forward as the perfect way to test artificially intelligent machines out. At this point, even though artificial intelligence is well developed on the sensory front, to solely test a machine’s cognitive capabilities, text-based games stand out as the premium method to gauge exactly how a machine understands, analyzes, and then responds to a problem scenario. Microsoft’s TextWorld is an opensource Python-based framework that generates randomly designed unique text problem scenarios which allows tech developers to observe how their AI devices learn and retain their learning by making use of language for understanding and answering as well as employing prompt decision making in a domain of set conditions and possibilities of turns. This project is brought forward by Microsoft’s acquired FATE AI lab in Montreal and the product has since been made available on their site for free download as of the 12th of July, 2018.

TextWorld Game Simulation Prompt. Microsoft

As TextWorld stands now, the problems take place in a house. This is restricted to allow the AI to gain familiarity with its environment so that its attempt at subsequent problems can also reflect upon its retention of what was learned in the solutions to previous ones. Most of the problems revolve around basic in-home tasks such as transporting objects around, interacting with different parts of the house, and performing day to day chores. This simulation game acts as a playground to test and develop AI for learning retention and effective decision making. In the case of this game, the two components of game generator and game engine work hand in hand. The former creates bounds for the game to develop in. These bounds include the number of rooms, stories, objects, and objectives that create the setting that the game takes place in and determine what is needed to successfully defeat the game. The game engine then uses these pre-set conditions to create game specific scenarios for the actual game play that follow a single command parsing mechanism in binary modules that allow for the game to move forward once the correct response is initiated or move backwards if a consequence of the wrong response is to be carried out. The implication of consequences demands that the machine playing the game not only decide upon the correct set of response commands but also the correct sequence and the correct timing in order to smoothly move forward through the puzzles. Extra players are not yet a part of the game

As many AI developers have tested out the game, the concern remains that some of the scenarios and commands in the game are rather incomplete for the machine to make decisions about. Some scenarios in the game are also considered to be “too easy’ as well but for the purpose that they serve, as it stands, the game does engage the cognition centers of artificial intelligence run machines. The tech experts at Microsoft’s base in Montreal are enthused to see the implications of their developed game and many AI developers are rushing about to wrap up their products for testing in an open summit of this year’s IEEE Conference on Computation Intelligence and Games (CIG) on the 20th of July this year. The summit will feature a competition that tests AI machines against this game and this is the best opportunity for individual startups and private developers to test their products against a set industrial standard.

Presentation at the IEEE Conference on Computational Intelligence and Games. IEEE CIG

Zainab Imran

Zainab Imran is an Electrical Engineer with a keen eye for innovative designs and computer system enhancements. Her knowledge and professional experience in dealing with the world of technology fuel her passion to thoroughly seek out and celebrate the most trendy and ingenious developments with a positively critical approach in utter awe of what more is to come.