Planning strategies in the real world takes imagination.

A team at DeepMind has developed an algorithm that can figure out correct strategies efficiently in novel and complex environments. Such environments are not nearly as straightforward as, say, the environment in a game of Go. Go has strict rules. The real world has its own rules but they do not constrain outcomes as the rules of Go do. Within limits, anything in the real world can happen. So, reliably training an algorithm for real world encounters is not feasible. Trial and error also suffers from the fact that some trialling can be catastrophic.

Humans deal with these problems by having a mass of experience in their heads already and by imagining alternative strategies. DeepMind’s imagination-augmented agent (I2A) mimics both capabilities. I2As learn about their environments in broad terms; they get the gist. This is efficient. They simulate alternative strategies and collect clues from the rejected ones that might come in useful in other simulations. They avoid engaging with the environment before they have imagined the consequences. And they still get the reward at the end of the day; it is reinforcement learning after all. But they think through the problem first. They sound quite human really.

Link to paper:

You might also like to browse other AI papers: