Unity, maker of the widely used game engine of the same name, just announced a totally new kind of video game: one designed from the ground up to be played by artificial intelligence rather than humans. This software will put AI through it paces and could act as a benchmark for future virtual beings.
Called Obstacle Tower, this software is created to judge the level of sophistication of an AI actor by measuring how well it can navigate up to 100 procedurally generated levels of continuously changing complexity and difficulty.
The idea is that his game will serve as a benchmarking standard to rate the ability of self-learning software. The AIs that do best in this virtual obstacle course can net their creators some serious cash, as Unity has set aside $100,000 as prizes for participants.
Speaking to The Verge, Unity’s vice president of AI and machine learning Danny Lange said that "What we really want to do here is create a tool for researchers to focus their work on and unite around and compare progress."
Video games are naturally training tools for AI because beating one requires the AI to use a varied skill-set: problem solving, critical thinking, pathfinding, identifying obstacles and dangers, adapting to unexpected situations and so on. Even older video games from the 80s, such as the 1984 Atari classic Montezuma’s Revenge, are still used to challenge AIs in the realm of puzzle solving.
If the androids and autonomous robots of science fiction are ever to become a reality, they will need to lean how to navigate the physical world using all the skills that can be learned through simulations such as Obstacle Tower. The trick is to have the AI to sense and 'see' the game world the same way a human would, to learn our visual language. What was so significant about Deepmind's victory over human pro Starcraft 2 players, for example, was that it didn't cheat, but instead used superior tactics with the same information a human would.
AI sophistication is making rapid strides thanks to innovations in machine learning techniques such reinforcement learning, in which an AI is given a reward mechanism to encourage improvement towards a particular behavior such as curiosity, which is of benefit to puzzle or exploration games.
But these AIs also need virtual training grounds to practice their skills. Obstacle Tower could well become the gold standard for researchers to tout for their AI, or be the first in a new genre of games specifically to be played by virtual agents.
Source: The Verge