Reinforcement learning (RL) is a powerful tool for sorting out decision-making problems, and it’s being used to train A.I. systems to achieve superhuman capabilities.

Inside of a computer simulation, a system tries, fails, learns, experiments and then tries again, in rapid succession, altering its future attempts each time. It’s because of RL that AlphaGo, a computer developed by DeepMind (part of Alphabet), learned how to beat the greatest Go players in the world. One problem with RL: agents have difficulty when they don’t have enough supervision, or when their objective is to run scenarios for a very long time horizon.

In 2020 and beyond, researchers will try to solve those problems using hierarchical reinforcement learning, which discovers high-level actions and methodically works through learning challenges to master new tasks at speeds we humans can’t imagine. This is important for non-techies, too: RL will improve the “intelligence” in our A.I. systems, helping cars learn to drive in unusual conditions and helping military drones perform complicated maneuvers that have never been attempted before in the physical world.