A New Kind of AI – Augmented Random Search

Recently, a phenomenal topic has emerged among AI researchers. A new kind of AI has been invented!
The algorithm is so simple that you don't need any complex framework like TensorFlow to implement it; you can try it out in just a few dozen lines of code directly in Python or a language you're comfortable with. What's fascinating about it is that it is 100 times faster and more efficient than traditional algorithms like deep learning or reinforcement learning. What is this algorithm that outperforms the most powerful algorithms from the labs of Google and Facebook in benchmark tests? Is it really true, or just an inflated bubble?

The algorithm is called ARS – Augmented Random Search. Although the impressive results of this test were published in research on 19 March 2018, I believe it is the same algorithm known since 2017 as the Evolution Strategies Algorithm. It appears that its strength lies primarily in tasks where you give the AI model a little figure with a few virtual muscles, and the AI tries to figure out how to move the virtual being from point A to point B as quickly as possible. Here, ARS truly shows interesting results.

How does ARS differ from traditional AI? Primarily in three points.

1. ARS does not adjust the weights on the neurons after every action, as is done in reinforcement learning, but only at the end of each episode.
2. ARS does not use the classic Gradient Descent algorithm to update weights, but rather a very computationally simple calculation (based on adding and subtracting random numbers).
3. ARS does not employ a deep network architecture (like deep learning) but only a very primitive single-layer perceptron.

Based on these comparisons, I believe that the strong points of ARS will be speed and simplicity, albeit at the cost of reduced universality and intelligence.

Famous research: https://arxiv.org/pdf/1803.07055.pdf
Earlier research: https://arxiv.org/pdf/1703.03864.pdf
Course, where it is simply explained: https://www.udemy.com/artificial-intelligence-ars/
Video: https://www.youtube.com/watch?v=YC-Ll76BpM8