Reinforcement learning: the next big AI trend?
Deep learning has been the star of the AI show for a while, but there are several up-and-coming areas that deserve the spotlight. One of these is reinforcement learning. Industry leaders like OpenAI and Google’s DeepMind have made major breakthroughs in this area already.
Surprisingly enough, reinforcement learning isn’t new – it’s a concept that’s been known to psychologists since the early 50s. Put simply, it means a behaviour becomes more or less frequent depending on its direct consequences. That goes for humans, animals – and AI. In AI it’s often known as deep reinforcement learning because deep learning is used to leverage it.
Reinforcement learning involves an agent (the AI), an action, and a reward. Just like someone playing a video game, the AI tries out actions over and over, observes the consequences, and learns to adjust its behaviour to get the maximum rewards.
In fact, video games have been some of the most intriguing use cases for this technique. For instance, DeepMind’s AlphaGo used reinforcement learning to teach itself to play Go – a game with more potential moves than there are atoms in the universe – well enough to beat world champion Lee Sedol.
Video games and movies could also be transformed by AI characters powered by reinforcement learning, who adapt to the player or viewer to cocreate elaborate new storylines rather than sticking to fixed dialogue and interaction.
Reinforcement learning also has fun applications in the real world: OpenAI used it to create a robotic arm that could solve the Rubik’s cube. But it has more serious potential uses too – like finding the best tax policies by looking at how different policies affect economic agents’ behaviour.
We could also see AI doctors using reinforcement learning to learn how to treat patients. Of course, we can’t allow them to experiment on people, but we can let them loose on historic patient data, where they can come up with hypotheses and then ‘test’ them by searching the data.
Reinforcement learning is still a newcomer, but given the breakthroughs it’s already had, we can expect to hear a lot more about it and to see it disrupt a wide range of industries.