University of Arizona and Microsoft AI Research Presents ‘TextWorldExpress’: A High-Performance Text-Game Simulator That Boosts Simulation Upto One Million Steps Per Second

Even though they only have a virtual, not a physical, embodiment, agents that are graphically portrayed with a body, such as a human or a cartoon animal, are also referred to as embodied agents. The development of artificial agents that can function and reason in embodied contexts has long been one of the main objectives of artificial intelligence.

Text games, or settings created solely in natural language, have recently gained popularity as an alternative research approach for embodied agent research. This is because they have a lower entry barrier than 3D games and can easily mimic complicated tasks at a high level. Text-based video games simulate an agent moving around an environment and writing down what they see. Similar to how people engage with their surroundings in 3D environments, agents interact with the world via abstracted high-level natural language commands.

However, text games require a range of common sense skills to be completed successfully, such as knowing how to read and follow instructions and comprehending world affordances. As a result, text games continue to be very difficult for agents to play, with the current state-of-the-art performance for classic interactive fiction games like Zork only 12%. The same may be said about clear step-by-step logic and interactivity.

In the new paper TextWorldExpress: Simulating Text Games at One Million Steps Per Second, a research team from the University of Arizona and Microsoft Research Montréal addresses this issue and suggests a high-performance text-game simulator that boosts throughput by roughly three orders of magnitude, reaching one million steps per second (SPS). 

The extremely optimized and profiled code used to create TextWorldExpress allows it to produce surroundings quickly while generating an extensive list of possible agent actions, greatly reducing simulation time. TextWorldExpress simulations can run purely on CPU cores, enabling it to run expensive multi-GPU nodes-free large-scale simulations.

Using tasks from three action spaces—CookingWorld, TextWorld Commonsense, and Coin Collector—the team compared the TextWorldExpress with three well-known benchmark environments: TextWorld, Jericho, and ScienceWorld. The findings demonstrate that TextWorldExpress is about three orders of magnitude quicker than existing simulators in simulating in online generation mode, with an average performance of 212k frames per second per thread. TextWorldExpress can go above one million steps per second on multi-core computers.

This Article is written as a research summary article by Marktechpost Staff based on the research paper 'TEXTWORLDEXPRESS: Simulating Text Games at One Million Steps Per Second'. All Credit For This Research Goes To Researchers on This Project. Check out the Preprint/Under review paper, github link and reference article.

Please Don't Forget To Join Our ML Subreddit