Google Unveils a New Version of AlphaGo Which Learns on Its Own

Wednesday, 18 October 2017 - 8:04PM
Technology
Artificial Intelligence
Wednesday, 18 October 2017 - 8:04PM
Google Unveils a New Version of AlphaGo Which Learns on Its Own
DeepMind
The time when humans are able to outsmart computers in logical games with clearly defined rules has long since ended, and we're used to the idea of being made obsolete when playing board games.

AI program AlphaGo's triumph over the world's best living Go player didn't come as a huge surprise to anyone who's been paying attention to the development of artificial intelligence software. That said, there is something extremely satisfying about learning that AlphaGo itself has now been proven woefully obsolete by its flashier, newer, more advanced younger sister.

AlphaGo Zero, the new iteration of artificial intelligence software from Google's DeepMind initiative, is not just better than AlphaGo - it's infinitely better. While AlphaGo was only able to beat Go player Lee Sedol four out of five times the two of them played last year, AlphaGo Zero has beaten AlphaGo at the game one hundred times in a row without fail.

What's more, AlphaGo Zero is capable of learning independently, without outside input. AlphaGo learned to play Go by observing thousands of human players before trying to learn to play the game itself in simulations. Instead of relying on puny human game strategies, AlphaGo Zero simply made moves at random, slowly learning strategies that worked and those that didn't over the course of countless games playing against slight variations on itself.

Thus, AlphaGo Zero learns far faster, and doesn't require a beginner's lesson from a human in order to start picking up a new skill.



For those who've long since lost interest in robots beating humans at strategy games (it does tend to happen a lot), there's thankfully a far wider application to AlphaGo's learning abilities. DeepMind engineers theorize that AlphaGo Zero can also be put to task with difficult scientific challenges, such as sorting through proteins within the human body to find more adequate medical breakthroughs, or analyzing weather patterns in order to better understand climate change.

Considering that robots are already replacing humans in a series of real world tasks, AlphaGo Zero will no doubt prove a useful tool at the disposal of many commercial companies looking to find ways to make their business run more efficiently. The genius of the AI program's design is that as it gets its knowledge from trial and error computations over time, it often finds ways to solve problems that would never have occurred to a human being. AlphaGo Zero's strength is in its ability to take in data from a vast number of simulations and apply it in a rational, logical way - something that we all struggle with at the best of times.

Google will no doubt find a way before long to make AlphaGo Zero effective in a lot of modern, commercially viable enterprises. That's got to be a big goal for the company, because if there's one task that AlphaGo Zero hasn't learned to perform yet, it's finding a way to offset its own development costs.

DeepMind cost Google around $150 million last year, and it looks like the invention of AlphaGo Zero may be a way for the company to make something out of the project that can generate income to make up for some of this loss. Google insists that ethics is a big issue with their creation, but it's worth assuming that making money is an even more pressing concern for the search engine giant.

No doubt nobody at Google is in any hurry to recoup the resources spent on DeepMind, though. The project is clearly working, and before long will almost certainly deliver the kind of moneymaking genius creation that Google is searching for.
Science
Science News
Technology
Artificial Intelligence
No