当前位置:网站首页>Is AI more fair than people in the distribution of wealth? Research on multiplayer game from deepmind
Is AI more fair than people in the distribution of wealth? Research on multiplayer game from deepmind
2022-07-07 17:10:00 【Zhiyuan community】
Yi Pavilion From the Aofei temple
qubits | official account QbitAI
DeepMind No chess this time , And don't play video games , Instead, I studied a multiplayer game .
The latest development of “Democratic AI”—— Learn human values through training , Then we can allocate resources fairly according to everyone's contribution .
To demonstrate this concept ,DeepMind Designed a simple investment game , from AI And human beings serve as referees respectively , Let players choose the preferred allocation rules ,Democratic AI Even got a higher support rate than human referees .
AI Referees are more popular than humans
When a group of people decide to concentrate their money on investment , How to distribute the income is a big problem that we must face .
A simple strategy is to distribute returns equally among investors , But this is likely to be unfair , Because some people contribute more than others .
The second option is , We can allocate according to the initial investment of each person , That sounds fair , But what if people start with different asset levels ?
If two people contribute the same amount , But one is a small part of their available funds , The other contributed all his assets , Should they get the same share of income ?
To meet this challenge ,DeepMind Created a simple multiplayer investment game .
Game involves 4 Players , Share in 10 round .
Each player will be allocated initial funds , In each round , Players can make choices according to their own wishes : keep , Or invest it in a common pool .
The investment will definitely pay off , But there is a risk —— Players don't know how the final revenue will be distributed .
besides , They were told , front 10 There is a referee in the round (A) Make allocation decisions , Then 10 round , By different judges (B) To take over .
At the end of the game , They will vote for A or B, To decide which referee you want to play with again .
And the revenue of this last game can be retained by the players themselves , This will enable players to choose the most impartial referee in their hearts more actively .
in fact , One of the judges performs according to the preset allocation rules , On the other side is by Democratic AI make designs of one's own .
When we study the voting of these players , We found that AI Designed rules are more popular than standard allocation rules .
meanwhile ,DeepMind Also invited a human referee , And introduce him to the rules 、 Let him try to achieve fair distribution to win votes , But the final vote showed , He still lost to Democratic AI.
Democratic AI Why can we win ?
stay DeepMind The latest is published in Nature Sub issue Nature Human Behaviour Papers , It records the researchers' understanding of Democratic AI Training process of .
First , They let 4000 Many human players participate in the game many times under different allocation rules , And vote to choose which distribution method you prefer .
These data are used for training AI To imitate human behavior in the game , Including the way players vote .
secondly , The researchers made these AI Players compete with each other in thousands of games , And another one. AI System basis AI Players' voting methods continue to adjust the redistribution rules .
therefore , At the end of the process ,AI Redistribution rules that are very close to fairness have been established :
First ,AI Choose to allocate according to the proportion of relative contribution rather than absolute contribution . It means , When reallocating funds ,AI We will consider the initial amount of each player and their willingness to invest .
secondly ,AI The system specially rewards players who contribute more generously , To encourage others to do the same . It is important to , AI can only discover these rules by maximizing human voting rate .
Can this method be extended to reality ?
although DeepMind The game test of has achieved brilliant results , But to transform this approach from a simple four person game to a large-scale economy , It is still a huge challenge , At present, it is uncertain how it will develop in the real world .
secondly , The researchers themselves found several potential problems .
Democratic One problem is that it may develop into “ Tyranny of the majority ”, This will lead to the persistence of existing discrimination or unfair patterns against minorities .
AI More work needs to be done to understand how design allows everyone's voice to be heard .
in addition , The researchers also raised people's concerns about AI The question of trust :
Whether people will trust by AI Designed mechanisms to replace humans ? If people know the identity of the referee , Will it affect the final voting result ?
If you want to Democratic AI Designed solutions are used to solve real-world dilemmas , This is crucial .
Reference link :
[1]https://www.deepmind.com/publications/human-centred-mechanism-design-with-democratic-ai
[2]https://www.nature.com/articles/s41562-022-01383-x
[3]https://singularityhub.com/2022/07/04/deepminds-new-ai-may-be-better-at-distributing-societys-resources-than-humans-are/
边栏推荐
- Skimage learning (1)
- LeetCode 312. 戳气球 每日一题
- 自定义View必备知识,Android研发岗必问30+道高级面试题
- LeetCode 213. 打家劫舍 II 每日一题
- 模块六
- Skimage learning (3) -- adapt the gray filter to RGB images, separate colors by immunohistochemical staining, and filter the maximum value of the region
- Smart logistics platform: make overseas warehouses smarter
- LeetCode 1626. 无矛盾的最佳球队 每日一题
- Arduino 控制的双足机器人
- SIGGRAPH 2022最佳技术论文奖重磅出炉!北大陈宝权团队获荣誉提名
猜你喜欢
SlashData开发者工具榜首等你而定!!!
Sator launched Web3 game "satorspace" and launched hoobi
[Seaborn] combination chart: facetgrid, jointgrid, pairgrid
Mrs offline data analysis: process OBS data through Flink job
Master this set of refined Android advanced interview questions analysis, oppoandroid interview questions
测试用例管理工具推荐
掌握这个提升路径,面试资料分享
Shallow understanding Net core routing
skimage学习(3)——Gamma 和 log对比度调整、直方图均衡、为灰度图像着色
skimage学习(1)
随机推荐
如何选择合适的自动化测试工具?
掌握这套精编Android高级面试题解析,oppoAndroid面试题
Skimage learning (3) -- gamma and log contrast adjustment, histogram equalization, coloring gray images
Skimage learning (3) -- adapt the gray filter to RGB images, separate colors by immunohistochemical staining, and filter the maximum value of the region
Mrs offline data analysis: process OBS data through Flink job
Sator推出Web3游戏“Satorspace” ,并上线Huobi
QT中自定义控件的创建到封装到工具栏过程(二):自定义控件封装到工具栏
[designmode] flyweight pattern
LeetCode 1654. 到家的最少跳跃次数 每日一题
应用在温度检测仪中的温度传感芯片
防火墙系统崩溃、文件丢失的修复方法,材料成本0元
【饭谈】如何设计好一款测试平台?
Arduino 控制的双足机器人
最新阿里P7技术体系,妈妈再也不用担心我找工作了
LeetCode 403. Frog crossing the river daily
Shallow understanding Net core routing
Master this set of refined Android advanced interview questions analysis, oppoandroid interview questions
Talk about the realization of authority control and transaction record function of SAP system
LeetCode 1626. 无矛盾的最佳球队 每日一题
99% 用户在 Power BI 云端报表常犯错误