A machine learning developer named Sebastian Schuchmann has recently developed a pair of neural networks. In the beginning, they were like siblings who only found common ground in society in a strange new world. He then forced her to fight for the entertainment of his YouTube subscribers.
The networks depicted in a video game world as a police cruiser and green jeep were created in Unity3D. In response to the video on his YouTube page, Shuchmann said:
The Unity people have a machine learning framework called ML Agents, which I use for most of my work. It implements algorithms like PPO (this is the same stuff that OpenAI uses). I'm in the process of moving more of my training to the cloud to reduce production time. However, this is associated with costs.
Each neural network has been designed to perform a specific task using reinforcement learning. In this case, the police car must prevent the jeep from crossing a white line. If it succeeds, it will be rewarded; if it fails, it will be punished. The opposite is true for the Jeep.
Probably your mind has not blown after you've seen that – the cars mostly get lost and circle with the occasional brief spit of movement that causes that makes sense. However, this is an excellent illustration of the operation of strengthening learning. Machines, under the current AI paradigm, find things out with brute force. At the beginning of the video, the cars are sitting there motionless, later they almost play the game.
Given enough time, these cars would likely display a superhuman tactic as they continue to optimize their networks to reach their destination. Some AI experts fear that such an "open" learning situation could lead to a doomsday scenario.
However, for now, this is just a cool way to gain insight into the thinking of machines with today's machine learning systems. More videos of Shuchmann can be found here.
More news and analysis from the world of AI can be found in our section Artificial Intelligence.