Interesting papers to read 4

Interesting papers to read 4

This paper was a breakthrough in AI for the reason that it clearly shows the possibility offered by the so called deep reinforcement learning starting from zero knowledge.

Playing Atari with Deep Reinforcement Learning

by Volodymyr Mnih

DeepMind's AI software went head-to-head against a professional video game tester in an impressive display of digital prowess.

The battleground? Classic video games. This research, published in the prestigious scientific journal Nature, sheds light on Google's AI triumph over human competition in various Atari 2600 games like Breakout, Video Pinball, and Space Invaders. Remarkably, the AI played at a level that closely rivaled human proficiency for the majority of the games.

DeepMind's AI employs two distinct techniques: deep learning and deep reinforcement learning. The groundbreaking innovation lies in merging deep learning with reinforcement learning—algorithms that enable software to evolve over time using reward-based systems.

By unifying these techniques, Google has engineered a "general-learning algorithm with applications in many tasks," according to Koray Kavukcuoglu, a Google researcher.

However, the applications aren't confined to just these domains. Industries like robotics and autonomous vehicles could stand to gain, particularly in areas dependent on computer vision.

Google's AI software, dubbed the "Deep Q network agent," achieved 75% of the professional tester's score across 29 out of 49 games tested. Notably, it excelled in Video Pinball. Deep Q exhibited proficiency in dynamic scenarios, such as ball bouncing in Breakout or engaging in video boxing. However, it struggled with long-term strategic planning, as evidenced in its poor performance in Montezuma's Revenge—earning a score of zero.

Despite these limitations, the notable breakthrough is linked to the fact that the AI algorithm does not use some prior knowledge about the Atari games so it gains the knowledge from zero like humans but has the advantage to learn at an impressive rate.

要查看或添加评论,请登录

Piero Giacomelli的更多文章

  • GuerrillaMaths 2#

    GuerrillaMaths 2#

    Today I am going to talk about another example of #GuerrillaMaths, on a subject that on the internet has often found…

  • GuerrillaMaths 1#

    GuerrillaMaths 1#

    The Cambridge Dictionary state that "Guerrilla Marketing" is defined as unusual methods of getting attention for a…

    1 条评论
  • Interesting papers to read 10

    Interesting papers to read 10

    Today we deal with a papers by one of the father of AI that create the word cybernetics. Some Moral and Technical…

  • Interesting papers to read 9

    Interesting papers to read 9

    According to today's paper, Give Your Ideas Some Legs: The Positive Effect of Walking on Creative Thinking By Marily…

  • Interesting papers to read 8

    Interesting papers to read 8

    Today I propose the reading of an interesting papers of 1959 by Giuseppe Cocconi (https://it.wikipedia.

  • interesting papers to read 7

    interesting papers to read 7

    Today I propose something link to art with a touch of computers in between "Paintings by Turner and Monet depict trends…

  • Interesting papers to read 6

    Interesting papers to read 6

    My suggestion for today is the paper "Deterministic nonperiodic flow" by Edward Lorenz https://journals.ametsoc.

  • Interesting papers to read 5

    Interesting papers to read 5

    Today I would like to share an article of 40 years ago that give the birth of the so called "Doomday argument". The…

  • Interesting papers to read 3

    Interesting papers to read 3

    Death is inevitable but some visionaries scientist are working to try to reverse it. It's not a 80' horror b-movie but…

  • Interesting papers to read 2

    Interesting papers to read 2

    There are tons of books out there on how to be a great leader. However, today I would like to propose a the scientific…

社区洞察

其他会员也浏览了