DeepMind Masters Games without the Rules, Just Rewards

MuZero. An intriguing  example, with the right data.

 DeepMind's AI Masters Games Without Even Being Taught the Rules

IEEE Spectrum by   Philip E. Ross

The MuZero artificial intelligence (AI) developed by Alphabet subsidiary DeepMind can master games without learning their rules beforehand. The system attempts first one action, then another, learning what the game rules permit, and concurrently noting the rewards proffered. MuZero then modifies its strategy until it finds a way to collect rewards more readily; this observational learning is perfect for any AI confronting problems that cannot be specified easily. MuZero outperforms previous DeepMind AIs in its economical use of data, because it models only environmental parameters that matter in decision-making. Once trained, MuZero processes so little for decision-making that its entire function might be managed on a smartphone. ... "