Once an AI model has specific knowledge or skills, it can easily compete with a human. However, the learning stage itself is the Achilles heel of this type of algorithm. As a result, problems that take a few seconds to solve can be insurmountable for machines. And even if you manage to deal with them, the amount of time and resources turns out to be disproportionate to the actual scale of the difficulties.
Read also: Artificial intelligence advises the prime minister. A European country has hired a robot
As noted by the authors of the post that is now Available as a preliminary printEncouraging the AI to read the user manual before starting the imposed task can speed up the learning process. This is what reinforcement learning looks like: You set a goal and then reward the AI for taking the necessary actions to achieve it.
Wanting to further improve the whole process, scientists from Carnegie Mellon University decided to help the algorithms learn faster. To do this, they combined it with a language model that is able to read user manuals. The effects were not long in coming: the AI learned to play the video game much faster than in the case of the model developed by DeepMind.
Artificial intelligence has been trained as part of what is called reinforcement learning
First, however, the language model had to be trained to be able to extract and summarize key information found in the official game manual. This data was later used to ask questions about this game. The answer, of course, was given by a trained mechanic. These were then used to generate additional rewards and fed into the reinforcement learning algorithm.
Read also: DuckDuckGo is going like a storm. DuckAssist is meant to be your assistant – smart but secure
Finally, it’s time to test. To evaluate their approach, the researchers tested it in a game known as skateboarding. When they compared the results achieved by other tools with those that could have been “thrown away” thanks to the new approach, their hands began to clap. Suffice it to say that before, artificial intelligence had to complete 80 billion methods to achieve performance comparable to human performance. However, scientists managed to reduce this number to 13 million. So we’re talking about a 6,000 times better result. Other more complex products, such as the popular Minecraft, await next.
“Infuriatingly humble musicaholic. Problem solver. Reader. Hardcore writer. Alcohol evangelist.”
Hawking’s paradox has a solution? Scientists on the trail of the mystery of black holes
Plant sounds. Who hears tomatoes?
Z 229-15 is something you won’t understand no matter how long you think. Astronomers gave up