Hellishly Hard New Game Is Specifically Designed to Confound AI
The game development company Unity announced a new contest for AI gamers on Monday, according to The Verge. The company put together a puzzle-solving 3D platformer game appropriately titled “Obstacle Tower.”
The game was designed to be hellishly difficult for AI, featuring precisely the types of obstacles and cognitive challenges that trip up game-playing AI systems — and the first teams to complete it will win shares of a $100,000 prize pool.
In contrast to machine learning, the more-common type of AI that learns to detect patterns from huge sets of data, reinforcement learning is more suited to undefined problems like the procedurally-generated levels of Obstacle Tower. That’s because a reinforcement learning algorithm learns to adopt new behaviors based on programmed rewards, sort of like a dog learning a trick.
“There’s a wide range of control problems, visual problems, and cognitive problems that you have to overcome to progress from level to level, and every level it gets harder,” Unity VP for AI and machine learning Danny Lange told The Verge. “We’ve had human players play and they can get to around level 15.”
Unity will release the first 25 levels of the game next month, so that so developers can train their algorithms on the sorts of puzzles that make up the core gameplay. Then on April 15, all 100 levels will drop and the full contest will begin.