Google's AI Dreamer Project: Look at the past to predict the future

Recently, researchers from Google's DeepMind project collaborated with the University of Toronto with an AI project called "Dreamer" - to test the effectiveness of reinforcement learning for with artificial intelligence. Dreamer's design focuses on relying on what the machine has learned and learned in the past to make choices for problems based on "speculation" about future results.

Dreamer project shows that this artificial intelligence is very effective in processing data.

According to the researchers, the initial results of the Dreamer project show that this artificial intelligence is very effective in data processing, as well as has quite good performance in computing and processing - compared to with previous AI approaches.

Basically, Dreamer uses an operational model with a complex multi-part structure. A part of the system that encodes the observations as well as the actions of the system. Another part will predict the resulting states of the problem to be solved. The third part will give the evaluation value of the state of the problem, then based on this result to make a learning plan for the system - with the goal of predicting the steps to solve. problem solving. Dreamer's AI system will rely on existing inputs to "plan" and predict the system in advance of possible results, along with the "reward" of each result.

Exercises that AI Dreamer performs.

In their experiments, the researchers tested the Dreamer system on 20 exercises in the DeepMind Control Suite simulator. The result was that Dreamer took an average of 9 hours to achieve 106 steps, half the time that PlaNet - the predecessor of this project - needs to be able to achieve the same result.

According to the researchers, AI Dreamer is effective in using available models to accurately predict what they need to do to solve problems - even if they have never encountered them before. when. Not only that, Dreamer also proved effective for short-term plans, when comparing the results of 20 exercises with the old AI, Dreamer completed faster than 16/20 exercises, and achieved a tie at 4. remaining exercises.

If you are an AI researcher and curious about this project, the source code of Dreamer is currently posted publicly on the project's Github page.