Russia has introduced the world’s first public dataset – a database – for teaching artificial intelligence (AI) to make independent decisions. It consists of 100 billion data sets.

The new dataset is called XLand-MiniGrid and is designed to train “AI agents” – models who can make decisions. The solution was developed by a group of the AI Research Institute of Artificial Intelligence with the participation of the T-Bank AI laboratory.

“The dataset does not contain random data from the Internet with questions and answers, but trajectories of how the AI agent acted in our environment. The main goal is to use the dataset to obtain neural network models that would have “agency”. We have made it open so that scientists and practitioners can come up with new methods in AI research and expand the boundaries of model independence,” Vladislav Kurenkov, head of the Adaptive Agents group at AIRI, told RG.

An AI agent trained on the new dataset can explore the world around them to understand which actions lead to a good outcome and which to a bad one. Then he should stop and go into the phase of using “good actions”. The task of “researching and using” various actions on the part of AI is to independently find a learning mechanism that will give the best result in fewer steps.

Rossiyskaya Gazeta