Someone made an article about this. Great readRewards based on efficiency.
AI will solve the problem, but you have to guide it before you set it free, with either rewards or trained supervised. Otherwise it will solve a problem however it feels like solving a problem. An AI agent will jump off the game environment instead of running through a maze in order to beat it, for example.
Too clever for its own good! Google DeepMind researcher reveals how AI cheats at games | Daily Mail Online