Racing version of “doomsday man” staged speed and passion! Sony racing game professional players abused by AI

Man machine war is the most promising competition at present! Recently, alpha dogfight has defeated human pilots, but fortunately, in racing games, we human players have been very good. But recently, researchers from Switzerland have used deep reinforcement learning to play a racing car, and it is estimated that human beings will be eclipsed again < p > < p > in the final of DARPA “Alpha dog fight” challenge, the virtual aircraft controlled by AI algorithm surpassed human pilots and won by 5:0! < / P > < p > deep reinforcement learning has achieved good results in many decision-making fields, especially in games, where many games have reached or even exceeded the human level. < / P > < p > based on deep reinforcement learning, alphago zero developed by deepmind completely abused human beings in go without using any human go data; dota five developed by openai reached the top level of human players in dota games; alphastar developed by deepmind also beat human professional players in StarCraft games. < / P > < p > whether on the real road or in the simulated environment, high-speed driving is a very challenging task, because it requires the driver to be fast, accurate and ruthless. At the same time, the physical performance of the car should be brought into full play. < / P > < p > until recently, deep reinforcement learning agents trained by researchers from the University of Zurich and the Federal Institute of technology in Zurich, Switzerland, broke this situation. < p > < p > the researchers chose GT sport, a popular racing game from Sony in 2017, which is popular among players with many models and dazzling tracks! < / P > < p > unlike in the past, researchers use DRL to train a deep sensory motor strategy, which can be directly mapped from observation results to control commands. < / P > < p > first, the researchers defined a reward function for developing racing problems, and accordingly, a neural network strategy mapped input states to actions. < / P > < p > their goal is to build a neural network controller that can drive the car automatically without understanding the dynamics of the car and make it run as fast as possible without hitting the track wall. < / P > < p > to give you a clear understanding, the researchers invited TG, an expert in the field of Gran Turismo, to have an online competition with TA. < / P > < p > personal best lap time for more than 50000 human players from 70 countries, as well as built-in non player characters. < / P > < p > in the experiment, DRL beat the built-in NPC and exceeded the personal best lap speed of 50000 human players. < / P > < p > PS: it is generally believed that the NPC built-in in modern racing games cannot compete fairly with human beings. For example, compared with the fastest human driver, the current built-in NPC in GTS will lose a total of 11 seconds. In this reference setting, NPC is 83% slower than all human drivers. < / P > < p > the researchers believe that this is due to the agents’ ability to learn their own trajectories, which are similar in nature to those chosen by the best human athletes and can maintain a slightly higher speed when turning. < / P > < p > including training and assessment, the team completed the DRL test in less than 73 hours. Although their research is limited to timing tests on no other racing car on the track, the team plans to use more data efficient RL algorithms, such as meta-rl, for more “speed and passion.”. Developed a “plug and play” solar power generation scheme, and “5B” won a $12 million round a financing

Author: zmhuaxia