Google AI Learns To Play Atari Video Games: here's How They Did It

Google has developed artificial intelligence (AI) that can learn old Atari games just by watching them and fight back human opponents.

DeepMind Technologies, a British AI company that was acquired by Google in 2014, has created an AI gamer called Deep Q-Network that can play 49 video games from an Atari 2600. The AI gamer can not only play the classic Atari games but it can also beat top scores of skilled human players in more than 20 games. The AI gamer does not go through the game manual but instead watches humans play the game and learn from them.

"It really is the first algorithm that can match human performance across a wide range of challenging tasks," says Demis Hassabis, co-founder of DeepMind.

Hassabis also suggests that the AI gamer is very similar to humans when it comes to learning as it gets better and better incrementally before it becomes nearly perfect in playing a particular game.

The best performance of the Google AI gamer was on simple boxing and pinball games. However, the software also achieved high scores on Breakout, a classic Atari game that needs bouncing of a ball so that rows of bricks can be cleared. Hassabis suggests that the AI gamer also learned some tricks that professional Breakout human gamers use.

"To use reinforcement learning succesfully in situations approaching real-world complexity, however, agents are confronted with a difficult task: they must derive efficient representations of the environment from high-dimensional sensory inputs," says a study published about the experiment published in the journal Nature.

The study also remarked that while these learning processes are still being developed for computers, humans and many animals are adept at this type of learning by using a combination of brain processes adapted to analyzing sensory information as well as reinforcement learning.

Even though the AI gamer outperformed humans in some of the games, it also underperformed in some other classics such as Pac-Man.

Google has not revealed the purpose of developing the AI gamer. However, market observers suggest that the software can help the company improve upon its ads service.

AI is getting a lot of traction in the technology industry and Google is not the only company that is developing AI software. Baidu, which is the Chinese equivalent of Google has set up the Institute of Deep Learning and also hired experts to understand the potential of AI.

Social media website Facebook has also set up an Artificial Intelligence Research Lab and Yann LeCun, a deep learning expert will lead the division to unravel the prospect of AI.

In mid-2014, Twitter bought Madbits, a deep learning startup. Madbits utilizes deep learning methods for understanding an image's content.

AI has been used for many years to play games such as Backgammon. However, with the advancement in technology AI systems can manage and learn even bigger data sets. Google suggests that the next step will be to check the AI gamer's ability on 3D video games that were developed in the 1990s.

ⓒ 2024 TECHTIMES.com All rights reserved. Do not reproduce without permission.
Join the Discussion
Real Time Analytics