Among them, Nanyang Technological University Anbo (co-chairman of the conference), Tsinghua University Tang Pingzhong (co-chairman of the program committee), Nanjing University Levin (chairman of the Workshop), MSRA Qin Tao (chairman of the industrial forum), Nanjing University Yu Yang (sponsoring chairman) and Tianjin University Hao Jianye (propaganda chairman) are among the organizers, and Academician Yao Qizhi, winner of Turing Award, also serves as honorary chairman of the conference and will make a keynote speech.
The reason why this conference was promoted and established by Chinese people is that the research group of distributed artificial intelligence in China is growing day by day, and the application in industry is becoming more and more frequent. Domestic academic circles urgently need to hold a new high-level communication platform.
On October 13-15, 2019, the first international distributed AI conference was successfully held in Beijing National Convention Center. Lei Feng. com learned that there were 1 workshop, 2 tutorial, 3 specially invited keynote reports, 3 industrial keynote reports, 6 industrial specially invited reports and 48 oral speeches. As the first DAI conference, this conference received 37 papers, 13 of which were included (11 long papers and 3 short papers). In addition, in order to enrich the content of this conference, the conference organizers also invited 35 papers from AAAI, AAMAS, IJCAI, NeurIPS, ICML, ACL, KDD and other top conferences to make oral reports.
The meeting began on the 13th, and there was a tutorial in the morning and afternoon. Interestingly, the lecturer of tutorial in the morning was Noam Brown, a famous researcher of Depo AI, and the lecturer in the afternoon was Tian Yuandong, a very famous Chinese scholar in the field of Go AI.
Noam Brown is a research scientist in Facebook’s artificial intelligence lab. He has made remarkable achievements in computational game theory and machine learning. The most famous one is the development of Libratus, a two-person unlimited poker, in 2017, and Pluribus, a multi-player unlimited poker, in 2019, respectively, which defeated the top human players and made a sensation. Among them, Libratus was also listed as one of the 12 annual breakthrough scientific achievements by Science magazine, and Pluribus was on the cover of Science magazine.
In tutorial, Brown tells about the game AI with "imperfect information". In traditional game AI (such as chess and Go), all information is known to both sides of the game. What artificial intelligence needs to do is to search for the optimal solution as soon as possible. With the breakthrough of AlphaGo in Go, the perfect information game AI has reached its peak. But in the real world, many decisions are often not fully presented to the participants, such as poker, and we don’t know the information of the opponent’s cards. In tutorial, Brown first explains why the strategy (search algorithm) used in the past for perfect information games will collapse in imperfect information games, and then introduces new algorithms that focus on overcoming the corresponding challenges, especially covering virtual games and counterfactual regret minimization algorithms, as well as search technology for imperfect information games.
Tian Yuandong comprehensively reviewed the methods, tools, applications and other aspects of game AI from the perspective of reinforcement learning. Tian Yuandong is the head of Facebook in the field of Go research. At the beginning of this year, Tian Yuandong completely opened the ELF OpenGo pre-training model and code developed based on AlphaZero research work, which became the first open source of Go AI, greatly promoting the extensive research and application of Go AI.
In tutorial, Tian Yuandong extensively reviewed the latest reinforcement learning methods (A3C, APE-X, R2D2, SAC, self-learning, etc.) and their usage in games and other applications, and put forward a brand-new reinforcement learning tool ReLA. Tian Yuandong said that ReLAx can use native vector support in PyTorch C++ API, has efficient batch processing function, and can perform parallel network forwarding, compared with the ELF they used for open source before. Tian Yuandong then explained the application of ReLA through more detailed application cases, which is worthy of careful study by the researchers of game AI development.
Another highlight of this conference is the lineup of specially invited keynote speeches composed of Yao Qizhi, Victor R. Lesser and Liu Tieyan.
Yao Qizhi is a world-famous computer scientist and winner of Turing Prize in 2000. Its main research direction is computational theory and its application in cryptography and quantum computing. In 1995, it proposed a distributed quantum computing model, which later became the basis of distributed quantum algorithm and quantum communication protocol security. In this conference, Yao Qizhi gave a speech entitled "Fintech: a meeting of minds between computer science and economics" as the first special guest speaker.
Yao Qizhi thinks that financial technology can be regarded as a fusion of economics and computer science in the digital age. The main technical basis of financial technology includes reliable distributed computing and cryptography of computer science and effective financial activity mechanism in finance. In the report, Yao Qizhi discussed some latest work in the field of auction and blockchain from the above perspective. For example, can you get more benefits from the auction that bidders are willing to pay? Will bidders have more income when they are more risk-tolerant than others? He then introduced some latest results about blockchain costs. He believes that these results are helpful to reveal the structural problems in economics, and the answers to these problems were not obvious in the past. The following is Yao Qizhi’s summary at the end of his speech:
Victor R. Lesser is one of the founders in the field of multi-agent systems. His key research areas include the control and organization of complex AI systems, and he has made outstanding contributions in the fields of multi-agent and blackboard system. He was the founding chairman of AAAI Fellow, IEEE Fellow, International Conference on Multi-Agent Systems (ICMAS) and International Association of Agents and Multi-Agent Systems (IFAAMAS). In 2007, in recognition of his outstanding contributions in the field of multi-agent systems, IFAAMAS also set up the "Victor Lesser Outstanding Paper Award" named after him. In addition, he also won the IJCAI "Outstanding Research Award" and other important awards in 2009. As the founder of multi-agent system, Lesser’s report is entitled "Reflections on Dai History and Coordination Technology", which comprehensively reviews the research history of distributed AI and collaboration technology.
Lesser recalled that in the late 1970s, a new field of distributed artificial intelligence began to rise, including distributed problem solving, planning, organizational control, negotiation and cooperation. The first seminar on distributed artificial intelligence was held at MIT in 1980, when only 22 people participated. It was not until 1995 that the first international conference ICMAS (International Conference on Multi-Agent Systems) was held. Then the research of distributed AI gradually flourished. Lesser lists people’s views on distributed AI in 1980s, and points out that at that time, "agents’ views can be uncountain, incomplete and out-of-date", although the bandwidth of users has increased exponentially after so many years, agents still have to deal with limited and outdated network state views. Then Lesser talked about his personal role in cooperation in multi-agent. And stressed that "collaboration was, is and will remain an important and challenging issue in distributed AI".
Liu Tieyan is the vice president of Microsoft Research Asia, IEEE Fellow. As a well-known expert in the field of machine learning and information retrieval, he has also made great achievements in deep learning, reinforcement learning and distributed machine learning in recent years. It is worth mentioning that the Microsoft team he led not long ago brought a breakthrough achievement to the AI field-the world’s strongest mahjong AI "SuphX" was promoted to ten stages in the Japanese online mahjong competition platform "Tianfeng". However, in this report, he did not introduce the mahjong AI they developed, but introduced how Microsoft Research Asia used artificial intelligence to help traditional enterprises carry out digital transformation with the theme of Towers AI-Powered Industrial Digital Transformation.
Liu Tieyan first introduced their cooperation with mutual fund company AMC and insurance company China Taiping in AI investment. According to Liu Tieyan, the AI investment model they developed achieved excessive returns and very good risk control. Then he introduced the cooperation between Microsoft Research Asia and OOCL, the world’s largest ocean transportation company. The technology of "competitive reinforcement learning" invented by Microsoft Research Asia solved their problem of relocating empty containers and greatly reduced the operating cost of OOCL. Liu Tieyan said that with the development of AI technology, more and more industries will undergo digital transformation, and AI scientists and domain experts should fully cooperate to jointly promote the progress of the world.
In addition to the special keynote speeches of the above three conferences, the conference also invited Qi Yuan, vice president of Ant Financial, Jiang Daxin, chief scientist of Microsoft Asia Software Technology Center, and Ye Jieping, vice president of Didi Chuxing, to share the application of distributed AI in their respective industries:
Qi Yuan, Vice President of Ant Financial Services
Report topic: multi-agent machine learning for all-inclusive finance
Jiang Daxin, Chief Scientist of Software Technology Center of Microsoft Research Asia.
Topic of the report: Question Answering in Bing
Ye Jieping, Vice President of Didi Chuxing
Report topic: AI for Transportation
This year’s DAI conference announced the Best Paper Award and the Best Paper Honorary Nomination Award at the dinner on the 14th.
The winners of this year’s best thesis are Weixun Wang, Jianye Hao, Yixi Wang and Matthew E. Taylor. Among them, Weixun Wang, Jianye Hao and Yixi Wang are from Tianjin University, and Matthew E. Taylor is from the University of Washington.
Title: achieving cooperation through deep multi-agent compensation learning in sequential prism’s dilemmas
Authors: Weixun Wang, Jianyi Hao, Yixiwang, Matthew E. Taylor.
Address: http://www.adai.ai/dai/paper/29.pdf.
This paper considers a multi-agent interaction problem. Considering that the real world is more of a multi-agent problem, the traditional "perception" method is not enough, and multi-agent research is needed to better simulate the real situation. Iterative prisoner’s dilemma has guided the research on social dilemma for many years. However, this problem is only divided into two kinds of atomic behaviors: cooperation and confrontation. In the real-world prisoner’s dilemma, these choices may be extended, and different strategies may bring a series of chain reactions, thus affecting the degree of cooperation. In this paper, the researcher puts forward a problem called sequential prisoner’s dilemma (SPD) in order to better capture the above features.
In this paper, the author puts forward a deep multi-agent reinforcement learning method, which can explore the evolution process of mutual cooperation in SPD problems. The researcher’s method is divided into two steps: the first step is the offline process, integrating strategies through different cooperation levels, and then training a cooperation level detection network. The second step is the online process, in which an agent gradually adjusts and selects its own strategy based on the detected cooperation level of the other party. The researchers believe that their proposed method can be demonstrated in two typical two-dimensional SPD problems: the "apple-pear" problem and the "fruit collection" problem. The experimental results show that the proposed method can prevent the agent from being exploited by predatory opponents, and at the same time, it can reach cooperation with those who are willing to cooperate.
The honorary nomination award for the best paper in this DAI conference was awarded to Song Zuo of Google Research, who interpreted myerson’s optimal auction theory from a novel perspective of linear programming.
Title: Rediscovery of Myers on’s auction via primal-dual analysis.
Author: Song Zuo
Address: http://www.adai.ai/dai/paper/36.pdf.
The theory of optimal auction was put forward by Myerson in 1981. This theory tries to solve the problem of how to design a system to stimulate the participants in economic activities to the greatest extent, that is, the design of optimal contracts, given the distribution of information. In 2007, Myerson won the Nobel Prize in Economics for this theory.
In this article, the author rediscovered Myerson’s optimal auction with completely different methods (linear transformation and original dual analysis). Specifically, he also considered the realization of Bayesian (Bayesian incentive compatibility+Bayesian individual rationality) and dominance strategy (dominance strategy incentive+ex post personal proportion), in which all buyers have additive valuation and quasi-linear utilities and all valuations are distributed with limited support. When the buyer’s value is one-dimensional and distributed independently, it can be directly proved that Bayes’ dual goal of implementing linear program does not exceed that of the dominant strategy. In other words, the optimal income under the implementation of Bayesian and dominant strategies is the same.
According to the author’s observation, if the dual plan is interpreted as the maximum virtual welfare, Myerson’s optimal auction can be directly interpreted as a "dominant strategy" linear plan. In addition, the author also describes the necessary and sufficient conditions for BIC = DSIC, that is, the optimal payoff of Bayesian realization is equal to the optimal payoff of dominant strategy realization (BRev = DRev). The condition is that the optimal DSIC income DR-EV (the virtual value of one buyer is independent of the valuation of other buyers) can be obtained only when the DSIC and post-IR virtual welfare independent of the virtual value function reach the maximum.
Lei Feng. com reported.
Related articles:
Yao Qizhi and Liu Tieyan gave lectures, Noam Brown and Tian Yuandong gave lectures, and the agenda of the first international distributed AI academic conference was fully announced.
China’s academic strength should not be underestimated. The first international conference on distributed artificial intelligence led by Chinese will be held in Beijing.
The original article of Leifeng. com is prohibited from being reproduced without authorization. See the instructions for reprinting for details.