Google’s AI platform has managed to learn to play video games. With not much of a practice, she was able to win in some of them.
This comes as a part of the development of Google’s DeepMind, which aims to examine the possibilities of artificial intellect. For the purpose of the test a 49 classic Atari 2600 games were used, including Breakout, River Raid, Boxing and other. The result was more than impressive. DeepMind managed to provide equivalent to a human player in half of played games.
Not so good was the performance of the Deep Q-Network (DQN) in the game Space Invaders. The reason is that in order to win, the algorithm has to develop a long-term strategy, which is still difficult.
In more simple games, however, DQN can learn how to beat them with practice.
“Surprisingly, DQN can immediately begin to play all the games without difficulty. I had to give her only information about what is happening on the screen, the available actions in the game and how to make points,” said Harsha Kumara which is one of the authors of the project.
Platform, which is based on the principle of the human nervous system, has begun to remember what happened in previous game attempts. She has used the information available, to change her approach in subsequent experiments until finally managed to find a way to start winning.
For now, the AI is still in a phase of training and development. However, Google wants one day to embed their services. The company hopes to make it so that one day the users can tell Google Now to plans a mountain trip through Europe, and the voice assistant to instantly offer solution.