《人工智能造福人类的那一面(48页).pdf》由会员分享,可在线阅读,更多相关《人工智能造福人类的那一面(48页).pdf(48页珍藏版)》请在三个皮匠报告上搜索。
1、AI for Social Good 人工智能造福人类的那一面 Prof.Toby Walsh UNSW Sydney|Data61|TU Berlin 澳大利亚新南威尔士大学教授、AAAI执行委员会成员 From food banks to organ banks Poverty 23 million people in Australia 2.2 million in poverty 11%children 25%pensioners Over 100,000 homeless FoodBank Local Social startup Winners of Microsoft Imagi
2、ne Cup(Australia)Finalists worldwide Using technology To reduce friction for FoodBank Australia(and other NGOs)Collecting&distributing food Fair division To different charities Pickup&delivery problem Induced traveling salesperson problem Online fair divison Goods arrive one by one Agents see items
3、and bid Only 0/1 utilities Special features Online Repeated Combinatorial Storage Expiry dates Unequal entitlements.Like mechanism Agents bid for any item with non-zero utility Item allocated uniformly at random to any bidder Balanced Like mechanism Agents bid for any item with non-zero utility Item
4、 allocated uniformly at random to bidder with fewest items Normative properties THM Like is strategy proof.THM Balanced Like is strategy proof for 2 agents but not for 3.Normative properties THM Both Like and Balanced Like are envy free ex ante THM Balanced Like is envy free up to one item ex post.D
5、eceased organ donation In 1989,average organ was 32 years old.In 2014,average organ was 46 years old.Fair division of organs Online Blood types Age groups Geographical regions.Blood types Supply tracks population Demand different Blood type B at disadvantage No help that O are universal donors Organ
6、&patient quality Kidney Donor Profile Index(KDPI)age of donor,.Expected Post Transplant Survival(EPTS)age of patient,.BOX mechanism Lexicographical preferences Blood/tissue type KDPI and EPTS Time on waiting list,.If KDPImax then 0,exit If KDPI=50 and EPTSEPTS-50 then+3000000,goto 2 If EPTS-50=KDPI=
7、EPTS-25 then+200000,goto 2 If EPTS-75=KDPI=EPTS-50 then+100000,goto 2.BOX mechanism Lexicographical preferences Blood/tissue type KDPI and EPTS Time on waiting list,.BOX mechanism Lexicographical preferences Blood/tissue type KDPI and EPTS Time on waiting list,.MIN mechanism Amongst compatible blood
8、/tissue type minimize|KDPI-EPTS|tie break by time on waiting list,.Why MIN?This is two-sided matching with identical preferences Patient wants organ with smallest KDPI Organ wants patient with smallest EPTS Stable organ matching Two-sided matching with identical preferences Unique stable matching it
9、h ranked patient with ith ranked organ MIN=stable matching Two-sided matching with identical preferences Unique stable matching ith ranked patient with ith ranked organ But online so what is ranking?MIN=stable matching Two-sided matching with identical preferences Unique stable matching Matching wit
10、h|EPTS-KDPI|minimized Suppose each is population percentile(which they are!)Formal model At each time step some patients arrive OR some patients depart OR some organs arrive Formal model Organs are matched on arrival each organ has KDPI each patient has EPTS Normative properties THM MIN is organ mon
11、otonic THM MIN is patient monotonic Normative properties THM No mechanism satisfies participation THM Only strategy proof mechanisms are random Waiting time Total waiting time constant MIN distributes this evenly Waiting time Total waiting time constant BOX does not From food to organ banks Both onl
12、ine fair division problems Special features we can exploit(like identical preferences)Normative analysis useful Tradeoff between fairness&efficiency From food to organ banks Join me(&others)in doing AI for social good Computational sustainability Security games AI&Education AI&Health .For more on“AI for Social Good”For more on“AI for Social Good”