A new research study from Carnegie Mellon University has found that AIs aren’t typically efficient learners, as it’ll take them a decent amount of time and data to solve problems that a human can get done instantly.
The researchers found that getting AI to read instructions before attempting a task could elevate their learning skills. Referred to as ‘reinforcement learning’, the technique involves setting a goal and rewarding the AI for taking actions toward completing that goal. The team figured out a way to help reinforcement learning algorithms learn a lot faster than usual by merging them with a language model that can actually read instruction manuals.
The result was that the team was successful in teaching an AI to play the classic Atari video game, Skiing, up to 6,000 times faster.
Yue Wu, the research lead says, “Our work is the first to demonstrate the possibility of a fully-automated reinforcement learning framework to benefit from an instruction manual for a widely studied game.
“We have been conducting experiments on other more complicated games like Minecraft, and have seen promising results. We believe our approach should apply to more complex problems.”
The research team commenced training the language model to extract and summarise key information from the game’s official instruction manual and used the data to ask questions about the game to a pre-trained language model.
Afterwards, the results were used to create more rewards for the reinforcement algorithm and then placed into an established reinforcement learning algorithm to help the AI learn Atari a lot faster.
In accessing their approach, the team ran the test on Skiing 6000 where the AI had to go through 80 billion frames of the game to achieve comparable performance to a human and found the new approach required 13 million frames to understand the game.