Making a neural network run through simple video games is an ideal way to test the effectiveness of its training due to the simple ability to evaluate the results of the passage. Developed in 2012 by DeepMind (part of the Alphabet holding), the benchmark of 57 iconic Atari 2600 games has become a litmus test for testing the capabilities of self-learning systems. And here is Agent57, an advanced RL agent (Reinforcement Learning) DeepMind, the other day a huge leap from previous systems and was the first AI iteration to outperform the baseline of a human player.
AI Agent57 takes into account the experience of the company's previous systems and combines algorithms for efficient environmental exploration with meta-control. In particular, Agent57 has proven his superhuman skills in Pitfall, Montezuma's Revenge, Solaris and Skiing - games that have been a serious test for previous neural networks. According to research, Pitfall and Montezuma's Revenge force the AI ββto experiment more to achieve better results. Solaris and Skiing are difficult for neural networks because there are not many signs of success - AI does not know for a long time whether it is doing the right thing. DeepMind built on its old AI agents so that Agent57 could make better decisions regarding environmental exploration and game performance evaluation, as well as optimize the trade-off between short-term and long-term behavior in games like Skiing.
The results are impressive, but AI still has a long way to go. These systems can only handle one game at a time, which the developers say goes against human capabilities: "The true flexibility that comes so easily to the human brain is still beyond the reach of AI."
Source: 3dnews.ru