Unity, a leading maker of game development tools, today announced that it has created an unprecedented new type of video game, designed not to be used by humans but with artificial intelligence. The game is called Obstacle Towerand is software designed to assess the sophistication level of an AI agent by measuring the efficiency with which it can maneuver up to 100 levels that change and increase in difficulty in unpredictable ways. Each level is procedurally generated, so it changes every time the AI tries.
With Obstacle Tower and a $ 100,000 prize pool reserved for participants to claim as part of a contest, Unity hopes to provide IA researchers with a new benchmarking tool to evaluate self-learning software. "We wanted to give researchers something to actually work on, which would, to an extreme degree, challenge the skills of artificial intelligence systems that are currently in research and development around the world," said Danny Lange, vice president of AI and learning Unity machine. The Verge. "What we really want to do here is to create a tool for researchers to focus their work and to come together and compare progress."
Video games are among the most useful training tools for AI researchers because of the sheer amount of critical thinking, problem solving, and road planning needed to play and succeed in simple arcade titles. And for years, the only game that proved to be a particularly challenging obstacle for AI agents and therefore a solid benchmark for measuring AI skills was the 1984 Atari classic Revenge of Montezuma. The game, unlike most others in its time frame, provided few concrete feedback mechanisms for players. Instead, it rewarded the exploration and resolution of puzzles, as opposed to quick reflexes and precise aim. This made it especially difficult for researchers to train AI software to learn how he played the game.
However, AI agents are improving rapidly thanks to new approaches to machine learning, which Unity cites as a motivator to create Obstacle Tower. In November last year, the IA OpenAI lab published research showing how a unique approach to the technique known as reinforcement learning, in which an AI receives a reward mechanism and circulates through hundreds of years of accelerated game time, adapted to reward curiosity yielded record performance in Revenge of Montezuma.
Reinforcement learning is how DeepMind, Google's software, trained software to beat the best players in the world Go and, as of last week, until StarCraft II. But the technique is only traditionally effective in certain games where parameters can be tightly controlled and the goals set for AI agents are clear, concise, and free from possible distractions. For Revenge of MontezumaOpenAI encouraged its algorithm to exploit the game essentially by giving it a secret to find on the first level of the game, which encourages the agent to quickly cross the environment more than it otherwise would.
In case of Obstacle Tower, Unity is taking a similar approach to design, though it is adding procedurally generated levels that also change in physical design as the AI progresses. The game is essentially a modern approach Revenge of Montezuma. He mixes platforms and puzzle solutions that will have players looking for keys and avoiding enemies and spikes, so Lange says it should be an effective AI experience test in areas such as computer vision, virtual locomotion, and planning. It is also in 3D and in the third person, which will require AI agents to exercise a more sophisticated level of spatial awareness as they move through the levels.
"There is a wide range of control problems, visual problems and cognitive problems that you need to overcome in order to progress from level to level, and at each level it becomes more difficult," says Lange. "We had human players playing and they can reach level 15." Unity plans to do Obstacle Tower of open source, so that game developers and researchers can modify it as they see fit. You can also download and try out if you are interested in testing a game that was never meant to be played by a human being.
As part of the contest that will be held throughout the game, Unity says any participant can train an AI agent to scale the first 25 floors of the tower between February 11 and March 31. As of April 15, the full 100-story game will be available, with the winners announced on June 14. Unity says it will distribute cash prizes as well as travel coupons and credits to Google's Cloud Platform. It is unclear exactly how the competition will reward researchers, whether for overall performance or the first team, to develop an agent capable of over 100 floors, but Unity plans to release more information about the contest in the coming weeks.
The ultimate goal is for these types of new, specially developed software to help create smarter AI agents who can learn more complex skills at ever faster rates. Learning how to play a video game will not be applicable to most of the real world tasks that we will make a robot in the future. Chances are that we do not want robots that try or not vacuum their carpet or fry an egg a thousand times until they get it right. (Although we may well have the robot software practice these tasks using virtual simulation.) And just by training deep neural networks in massive data sets for a singular and restricted purpose – how to recognize objects in photos – can companies like Google to make advances in AI research on real resources that we can use in today's commercial products.
But by training AI to play video games without any instruction, researchers are getting a better understanding of how the mind solves problems and, more importantly, how it learns to solve new problems that have never been found before. These kinds of challenges, such as Unity's Obstacle Tower, could provide researchers with new avenues to continue working on these challenges, with an eventual milestone in creating what the industry calls general artificial intelligence, or AI software that can accomplish any task a human can.
"Many people think that artificial intelligence is about building better product recommendations on Amazon. But at the end of the day, you're actually solving more complex problems. It's about dealing with vision, control and other cognitive challenges, "says Lange. For Unity as a company, he adds that this type of work also helps establish your game development toolkit as a place where cutting-edge research can in future translate into industry breakthroughs. "We have a mission to democratize the development of games, but we also want to democratize AI. We want to make sure that many developers can get their hands on this. "