General, Research, Technology

Artificial Intelligence DeepMind Learns to Predict the Future Based on the Past

Many artificial intelligence systemsachieve goals based on ideas about the world based on past experience. They generalize them and shift them to new situations, which allows them to perform tasks even in conditions that they have not encountered before. And if, as a result of these actions, “the mission is completed,” the AI ​​receives a certain “reward”. This learning method is called reinforcement learning. However, this method of teaching artificial intelligence is also suitable for not only gaining new knowledge, but also predicting the future.

AI learns to predict the future. What's next?

How can AI predict the future?

Researchers from DeepMind andWhile studying the dynamics of the reward model, the University of Toronto set up AI in such a way that, based on past experience, it would not take any action, but predict it based on known similar cases. In this context, there is no talk of a full-fledged prediction of events. Now the model is only exploring the unobvious possibilities for the location of objects in the future, given their non-linear speed and winding path.

AI behavior model called Dreamer andIt consists of five basic elements. The first is a representative component that observes the object and determines its future actions. The second is a transition module, which foresees the possible state of the object, not knowing their initial properties. The monitoring module of the current state provides feedback signals, which is paired with the monitoring module of the final state, which allows you to compare the prediction with the result. Well, the crowning component of this complex system is the incentive component.

In a number of experiments, scientists testedDreamer on 20 tasks in the program DeepMind Control Suite. At first, they trained him with one Nvidia V100 graphic chip and ten processor cores. Each training cycle took 9 hours. In comparison, PlaNet, the predecessor of Dreamer, achieved similar results in 17 hours. As a result, Dreamer was able to use the studied models to draw conclusions about possible outcomes and did it fairly accurately. And what do you think, how accurately can AI predict the future? And is it possible to make such a forecast at all? Write about it in our chat in Telegram.

See also: Artificial Intelligence has learned to track down hackers

AI has long existed that usebig data for analyzing the situation and predicting outcomes, but they are geared specifically for working with, which is logical, a large amount of information that needs to be studied in advance. Dreamer has developed an algorithm that can predict the outcome no worse than specialized systems. Yes, at the moment, AI only predicts the position of randomly moving objects, but it also has a possible scope.

For example, models like Dreamer canuse for short term planning. That is, in theory, he will be able, based on the data from the work of the company department, to predict what they will achieve in the future with given performance indicators and a schedule. Then it will be possible to optimize the activities of people so that the “future has changed” and their performance has improved.