We have been analyzing the effects, in physical and mental health, of video games. Now it’s time to analyze the effects of this type of activity on an “organism” that is growing by leaps and bounds: artificial intelligence. The question is simple: What happens if an AI learns to play, for example, Minecraft? Is it smarter? The short answer is yes.
This is the conclusion of a team of Google Deepmind scientists, in collaboration with the University of Toronto, it reports that the algorithm of the Dreamer can learn to improve yourself when Learn to play Minecraft. And this is achieved in a short time.
In a study, Posted in NatureDanijar Hafner, Jurgis Pasukonis, Timothy Lillicrap and Jimmy Ba scheduled Google’s application to play Minecraft without training.
In recent years, computer scientists have learned a lot about how deep learning (Deep Learning) and train AI applications can be used to carry out seemingly intelligent activities, such as answering questions. It has also been shown that IA applications can train to play video games and obtain higher performance than humans. But we thought it was a redundant investigation: what can be obtained from a computer playing with another computer?
In this new study, the authors discovered that video games can produce progress, such as Help a AI application to learn to improve their skills in a short period of timewhich could provide robots with the necessary tools for good performance in the real world.
For this, the Dreamer was programmed to play Minecraft through the creation of a rewards system, specifically rewards for finding diamonds. With this approach, The application did not need to be taught to play; He only needed to know the parameters within which he could work, one of which included visualizing a virtual world.
Once the algorithm learned to play Minecraft, the researchers, a novelty was added: they only allowed him to play on a certain scenario for 30 minutes. At that time, the game was restarted with a completely new virtual universe. And what they discovered was that The algorithm quickly improved, reaching the expert level after playing only nine days.
The study suggests that the ability of the algorithm to imagine a future where all its objectives have been achieved allowed it to focus only on the tasks that led to the desired objective and then use them in each new virtual world that it found. This result could Use to help robots learn to achieve predefined goals in the real world.