× close
SoG training process. Actors collect data via audio self-play, and trainers run individually on a distributed network. (A) Each search generates a number of CVPN queries containing the input β. (B) The query is added to the query buffer and then solved by the solver, which investigates the situation more thoroughly through another call to GT-CFR. During resolution, new recursive queries may be added to the query buffer. Separately, the network is trained on mini-batches sampled from the replay buffer to predict the values computed by the (C) solver and the policy targets. credit: scientific progress (2023). DOI: 10.1126/sciadv.adg3256
A team of AI researchers from EquiLibre Technologies, Sony AI, Amii, and Midjourney is collaborating with Google’s DeepMind project to develop an AI system called Student of Games (SoG) that can learn how to beat and play humans at a variety of games. has been developed. something new.In their paper published in diary scientific progressthe group will discuss the new system and its features.
Over the past half century, computer scientists and engineers have developed the idea of machine learning and artificial intelligence, which uses human-generated data to train computer systems. This technology has applications in a variety of scenarios, one of which is playing board games and parlor games.
Teaching a computer how to play a board game and improving its ability to beat humans was a milestone of sorts in how far artificial intelligence has come. With this new study, the research team takes another step towards artificial general intelligence. This means that computers can perform tasks that are considered superhuman.
Until now, most computer systems built to play board games focused on a single game, such as chess. In building such systems, scientists have developed a kind of limited artificial intelligence. In this new effort, researchers have built an intelligent system that can play a variety of games that require different skill sets.
There are two main types of gameplay. One is a game with perfect knowledge, or a game with incomplete knowledge. The former is one in which both players have complete knowledge of the game, including the positions of all the pieces in the game. The latter is like poker, where only some of the relevant information is known to the individual player. SoG not only allows you to play both types of games; It can defeat a skilled human.
So far, it has defeated other AI systems and humans at Go, Chess, Scotland Yard, and Texas Hold’em Poker, and it could also excel at other types of games because it can learn how to play virtually on its own. Any game, the research team suggests, is expensive.
For more information:
Martin Schmid et al., Student of the Game: Integrated Learning Algorithms for Both Perfect and Imperfect Information Games, scientific progress (2023). DOI: 10.1126/sciadv.adg3256
© 2023 Science X Network