Sep 18, 2020 11:05:00

AI hits time beyond humans with 'Gran Turismo Sport'

Following

AlphaGo , which defeated the strongest Go player in Go, and AlphaStar , which overwhelmed professional gamers with the real-time strategy ' StarCraft 2 ', the PlayStation 4 game software ' Gran Turismo Sport ' also researched that the deep learning model overtook humans. The results have been announced.

[2008.07971] Super-Human Performance in Gran Turismo Sport Using Deep Reinforcement Learning
https://arxiv.org/abs/2008.07971

A deep learning model achieves super-human performance at Gran Turismo Sport
https://techxplore.com/news/2020-09-deep-super-human-gran-turismo-sport.html

Deep Learning Model Outperforms NPC, Player Records in Gran Turismo | Science Times
https://www.sciencetimes.com/articles/27310/20200915/deep-learning-model-outperforms-npc-player-records-gran-turismo.htm

On August 18, 2020, a research team from the Swiss Institute of Technology Zurich (ETHZ) and Sony AI Zürich learned about a deep learning model trained to play the driving simulator 'Gran Turismo Sport' released by Sony Interactive Entertainment. I published a paper.

According to a paper published on the pre-reviewed paper publication site arXiv, the deep learning model developed by ETHZ et al. Not only beat the CPU in Gran Turismo Sport, but also set the best record for more than 50,000 human players. It is said that he set the fastest time to break all.

You can see what the deep learning model that showed a run beyond human beings is by watching the following movie.

Super-Human Performance in Gran Turismo Sport Using Deep Reinforcement Learning --YouTube


The deep learning model that surpassed humans in 'Gran Turismo Sport' aimed to develop an artificial neural network (ANN) -based control system that can drive a race car independently without prior information. It was created by research.

In order to develop a deep learning model, the research team repeated simulations to map observation data and trained commands such as steering angle and throttle and brake adjustment.

In addition, the deep learning model is based on the play of anonymous player TG who said that he achieved the best result in the international competition of 'Gran Turismo Sport', and the strategy of turning a curve and maintaining speed with minimum handling etc. I learned.

As a result of comparing the records of the deep learning model trained in this way with the records of 50,000 people provided by

Polyphony Digital , the developer of 'Gran Turismo Sport', the times of the deep learning model are all the times included in the data set. It is said that it exceeded.

Mr. TG, who saw the deep learning model running away from human beings, said, 'The deep learning model runs very aggressively, but I think that this is a trick because it can take accurate actions. I can follow the same track, but 999 out of 1000 will rub the car against the wall and ruin my time, 'he said, praising the elaborate technique of the deep learning model. It is said that it was done.

ETHZ's Yunlong Song said, 'We are currently planning to develop a more general AI agent that can run on any course with different car types, and will hit other cars while driving at high speed. We will also work to solve the 'passing problem of autonomous vehicles', which is to overtake without fail. '