Blog: The Next World Champion Of Every Game Ever
Playing games has long been the pastime of almost every human being, if not everyone. Games are a way to escape the stress of the modern world, train your brain, and just have fun. Until you lose. At the most complicated game in the world. The one you have spent your whole life training to play. 100 games to 0.
That’s the plight of the now second-best Go player in the world, AlphaGo. If you think that’s a weird name, it’s because AlphaGo is actually a computer program. And the player that beat it is another computer program, called AlphaGo Zero.
These aren’t your typical number-crunching calculating 200 000 moves per second computer programs. They’re in their own league, and that’s because they are artificially intelligent computer programs. Yes, the AI created by humans is beating them at their own game.
If a computer program can beat the world champion of the most complicated game in the world 100 games to 0, imagine what it could do if I used it to play chess against my sister! But in order to do that, I need to find out how this program works.
AlphaGo and AlphaGo Zero operate on a principle known as Reinforcement Learning, or RL. RL is a subfield of AI, and pretty complicated.
What is RL
Reinforcement learning is best explained with an analogy (unless you are better at explaining things than me).
Imagine giving a baby the remote control to your TV.
It will probably just push random buttons. If nothing happens, because the TV is off, the baby will get bored.
If it figures out how to turn the TV on and switch channels, it will a) get better at channel-surfing and b) probably be more interested in the flashy thing.
This is basically how an RL algorithm works. Like the baby, it starts off with what its environment looks like. In this case, it’s a remote control and a TV.
It also gets a list of actions, like push the home button, or press the power button.
The algorithm will randomly experiment with them.
Until it finds a solution.
Here’s the fun part. You can assign point values to every action, positive or negative. Since the algorithm’s goal is to maximize points, it will avoid actions with negative points.
Reinforcement learning has huge applications. It’s already the best Go player, and it doesn’t stop there. Chess, checkers, tic-tac-toe, and even video games like snake, pong, every Atari game you could possibly think of, and way more!
Games aren’t the only applications of RL. There’s also the really cool application of robotics. Using RL lets robots learn how to do something by watching a video clip of it!
Backflips are one example, but they can do anything just by watching a video. Automation is a huge application for reinforcement learning.
Google is also using RL to reduce energy consumption! Hopefully while it was training there were no random spikes in energy consumption while training, but now, AI is being used to help save the environment!
Reinforcement Learning is a really cool application of AI. It already has amazing applications, like huge automation and energy consumption, and I’m really excited to see what comes next!