Elon Musk and his nonprofit research institute, OpenAI, are giving Google and DeepMind some tough competition in the area of artificial intelligence and machine learning. Just a few months after DeepMind showed off some killer AI moves, Elon Musk announced that they can do better than that in a much simpler way.
Ilya Sustskever, OpenAI's research director, announced during the EmTech Digital Conference hosted by MIT review in San Francisco that OpenAI has discovered a much easier alternative to machine learning, a method that will help researchers speed up the progress of machine learning.
Sustskever said that instead of using reinforcement learning techniques that are being used today, they used evolution strategies, an optimization technique that has been used for many years. Evolution strategy is inspired by natural selection where organisms learn to adapt to their environment. What the technique does is try different actions and identify which ones are the most effective.
Evolution strategy is easier because it is scalable in multiple settings. They tested the technique by building a software that learned to play 50 different Atari games and the artificial players were able to learn in just an hour. Reinforcement learning, on the other hand, takes one day before it can learn the same amount.
They also tested the technique in making a humanoid walk in a simulated environment. The humanoid was able to learn the actions in just 10 minutes instead of the normal 10 hours.
Sutskever said that despite the promise evolution strategies hold, a lot of research is still needed to find out what its strengths and limitations are. He added that once the technique can be fully utilized, it can create artificial general intelligence which has the ability to adapt to a wide range of complex scenarios.
