Interesting papers to read 4
Piero Giacomelli
CIO | Group IT Manager at Fidia Farmaceutici SpA | Packtpub Book Author | Manning Publication Reviewer | Passionate about math
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.
领英推荐
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.