cnBeta.COM_Chinese Industry Research Center AAAI 2022 Grand Opening: AlphaHoldem 获 Exceptional Lectures
Visit:
Ali Yun 2022 “Shang Yun purchase season” minimum 0.26
This is the work of the next military team, which is the development of another company's development of a scaled-type Dezhou machine AI process——AlphaHoldem. Introduction, the system's decision speed has increased by more than 1000 times the speed of DeepStack, and the results of the high-level Dezhou competition players have been announced.
Compared to the task of playing Yoongi, the Dezhou game is based on information that is not complete and is an incomplete intelligent research technique.
Dezhou is the most popular game in the world, because it originated in the early 20th century.
Dezhou's rules are to use one side card for each player, 52 cards in total, minimum 2 participants, maximum 22, general participants between 2 and 10.
At the beginning of the game, each player created two private cards to create their own "bottom card", followed by five public cards, followed by three, one, and one in the morning. There are two privately owned tiles, three shared tiles, the fourth public tile, and the fifth public tile, which can be played indefinitely after the game city. Translated tiles, transliterated tiles, river tiles. Figure 1 shows the completion of the game in Dezhou.
Picture 1: The first game of the game
Following the four-wheel push note, if you can't win, you can't win the game. Figure 2 has been released.
Picture 2: Dezhou tiles
Dezhou's overcoming problems are very difficult to solve, two people are infinitely ordered over 10 to 161 steps; One-step decision-making depends on the first-step decision result, and at the same time, the opposite decision-making step produces an impact; ，It is necessary to fully consider various possible situations when making a decision at each step of the game.
Researchers, because of the Dezhou game game rules, are also very simple and world-wide, and specially adapted to work on a fictional, real-world environment-exploring related basic theory method and in-depth exploration of the core technology calculation method.
In recent years, an international researcher in Deok Province has made great strides in acquiring a large-scale imperfect information publication.
For example, before being a researcher at Kaman University, Wami National University, Naiki Meiryu University, AI program, DeepStack and Libratus, as well as people who have been in China for a long time, have been selected as a professional, followed by a follow-up card. The Pluribus designed by Naiki Meirung University also has 6 people who have won the championship in the Infinite Tournament.
However, the core idea behind Dezhou's current mainstream AI is the use of Counterfactual Regret Minimization (CFR), which is a close-to-balance strategy.
Concrete description, Abstraction technology [3][7] Compressive state and dynamic working space, from then on a small scale model, after that CFR calculation method to proceed to the expansion tree.
This small method is heavily dependent on the progress of human knowledge, and the demand for CFR calculations is constantly improving, and the demand for CFR calculations is constantly improving. . For example, DeepStack used 1.53 million CPU time and 13,000 GPU clock training final AI, the demand of the station level is 1 GPU progress 1000th CFR replacement process, average per operation calculation consumption time is 3 seconds. Libratus consumption is over 3 million CPU time generation initial strategy, every next decision search demand more than 4 seconds.
This large-scale calculation and resource consumption and resource consumption have made AI progress step-by-step research and development; Increased length. In addition, the abstract requires a large amount of knowledge of the area, and the inevitability of the inevitable loss of the area will be relevant to the decision.
This question has been sucked by many Chinese researchers, Institute of Automation This is one of the most interesting professor's missions. In December last year, another advanced research group signed up for Dezhou's board, and submitted a high-level, scaled two-person unlimited order for Dezhou's board AI process——AlphaHoldem.
Different based on CFR arithmetic Dezhou Coke AI, The structure submitted by the Chinese Academy of Sciences Research Group is based on the in-depth strong chemical calculation method (shown in Figure 4).
Figure 4: AI Learning Frame
Introduction to the root team, Actor-Critic learning framework adopted by AlphaHoldem, its import is card and action editing, after that, the special expedition provided by the same number of participants, and a kind of progressive depth-enhanced chemical calculation method Combined with a new type of self-learning learning method, regardless of the area of knowledge, you can directly receive information from the end of the study.
Instructed by others, the success of AlphaHoldem has been achieved due to its use of a highly effective state of the art. Optimal performance and speed, and a new type of Best-K self-exploration method that has been effective in solving the existing strategy problem during the expansion.
AlphaHoldem used 1 unit including 8 blocks GPU card service device, through three heavens self-exploration learning, battle Slumbot and DeepStack. Every time I make a decision, AlphaHoldem takes less than 3 seconds, and DeepStack speeds up over 1000 times. At the same time, AlphaHoldem announced the results of the 10,000th match against the four-ranked Dezhou National Team player, which has reached the level of the professional players.
Xingxian Liang, Researcher, Institute of Automation, Chinese Academy of Sciences, Doctoral Instructor, Special Youth Bone, Professor, University of Chinese Academy of Sciences, New Researcher, New Research Institute of Artificial Intelligence, Chinese Academy of Sciences. In 2012, Prof. Xing completed a Ph.D. in Engineering, Department of Computer Science and Technology, Qinghua University.
In addition, he is a senior member of the Institute of Electrical and Electronics Engineers (IEEE) of China, a special reviewer of the China Institute of Science and Technology, a senior member of the China Society for Computing Machinery (CCF), and a member of the Computing Committee.
Other main research areas are computer monitors and computers. Comprehensive international publications such as TPAMI, IJCV, AI, and other advanced international conferences such as ICCV, CVPR, AAAI, IJCAI published 100 papers, Google academic citations exceeded 10,000, published computer review 2nd edition, editorial copy 1 book in the field of deep learning, 1 book in the field of artificial intelligence.
Shenshu Qinghua University Arithmetic Department "Academic New Excellence", "Google Scholar", multi-level international and domestic conference best essay class, etc. International and domestic competition.
The current work is a multi-national key issue, researched and monitored, related to technology at the National Telecommunications Bureau. Usual.
In recent years, the major problem of intelligence sensing and solution related to intensive chemistry training has been researched. Dezhou's AI progress AlphaHoldem wins over 1000 times faster than the best public Dezhou's AI progress DeepStack. OpenHoldem, a large-scale imperfect information platform open to the academic world.
Executive essay:
Executive student thesis:
Excellent essay:
AlphaHoldem of the Chinese Academy of Science and Technology, with 5 works acquired AAAI 2022 “Outstanding thesis textbook”. how to separate
Authors: Bart Bogaerts, Stephan Gocht, Ciaran McCreesh, Jakob Nordström
Author Team: Jannik Peters
Authors: Thom S. Badings, Alessandro Abate, Nils Jansen, David Parker, Hasan A. Poonawala, Marielle Stoelinga
Authors: Jorge A. Baier, Carlos Hernández, Nicolás Rivera
Author team: Yanick Ouellet, Claude-Guy Quimper
Reference connection:
1. https://twitter.com/rao2z/status/1496866889921822721
2.https://mp.weixin.qq.com/s/OBRybZ-NwcNW-S9TCObaLA
![[Excel] How to paste images such as photos and diagrams](https://website-google-hk.oss-cn-hongkong.aliyuncs.com/drawing/article_results_9/2022/3/9/72539ecbf7413c05e4465b39ca06e8e0_0.jpeg "[Excel] How to paste images such as photos and diagrams")
