Dueling algorithms

Transcription

Dueling algorithms

Social Computing Models
Moshe Tennenholtz
Search
Search
• Say we know probability
distribution over pages
that the user desires.
movie reviews
prob. website
1.
18%
2.
16%
3.
15%
4.
13%
5.
12%
6.
9%
7.
8%
8.
8%
9.
7%
10.
6%
http://rottentomatoes.com
http://movies.yahoo.com
http://www.metacritic.com/movies
http://rogerebert.suntimes.com
http://www.movies.com/movie-reviews
http://www.moviefone.com/reviews
http://moviereviews.com
http://www.imdb.com
http://netflix.com/reviews
http://mrqe.com
http://movies.msn.com
• Trivial greedy algorithm
is optimal.
Which search engine
is better for me??
I’ll see which one ranks
(my favorite) higher?
(my favorie movie
From now on,
I’ll use Google.
movie reviews
http://www.kids-in-my-mind.com
http://pluggedin.com/movies.aspx
movie reviews
Perils of greed
Search
movie reviews
movie reviews
prob. website
18%
16%
15%
13%
12%
9%
8%
8%
7%
6%
http://www.imdb.com
http://mrqe.com
http://www.imdb.com
http://mrqe.com
Perils of predictability
Any deterministic algorithm can be badly beaten.
Search
movie reviews
movie reviews
prob. website
9%
18%
13%
15%
8%
9%
6%
8%
7%
12%
http://www.imdb.com
http://mrqe.com
http://www.imdb.com
http://mrqe.com
Optimization → Games
Ranking → Ranking duel
movie reviews
http://www.movies.com/moviereviews
1
𝑛
-beatable
Hiring problem duel
“Secretary alg.” is beatable
Routing
?
movie reviews
http://www.movies.com/moviereviews
http://www.moviefone.com/revie
ws
Greedy is
1−
Binary search duel
Solving duel is
→ Racecomputationally
harder
?
Binary search is
62.5%-beatable
Compression duel
Huffman coding is
66-75%-beatable
?
?
?
?
?
?
Shortest (expected) path is
1 − 𝜖 -beatable
TSP → Parking duel
TSP is
1 − 𝜖 -beatable
Routing
Which route should an agent take?
Taking the route that goes through s is α slower than taking
the route the goes through f, but service is splitted when
shared among agents.
Agent 1
Agent 2
f
s
Target
Pricing
Store
based
pricing
Pricing
Location based
differential
pricing
Display Advertising
Sponsored Search
Ranking Systems
On-line voting
• Voting (as preference aggregation) is popular in social networks
– For example, “likes” and “dislikes” in Facebook
– Usually, sequential voting: vote count shown as voting unfolds
– Other examples: voting in committees, US presidential primaries
On-line voting
• Voting (as preference aggregation) is popular in social networks
• Since vote is public, probably dislike voting against actual winner
– Facebook: disutility from disagreeing with most of my friends
– Committees: candidate may end up as a new faculty member
– Primaries: candidate may end up in a position of power
(Social) Trust Systems
Predicting and learning social structure
– How are friendships structured?
Zachary’s Karate Club – Social Network
from: “An Information Flow Model for Conflict and Fission in Small
Groups” W. W. Zachary, J. of Anthropological Research 33:4, 1977
• Conflict arose over price of lessons
• Eventually the club split to two clubs
• Structure -> Dynamics
• How do things spread in society?
– Epidemics, Information
“A History of Influenza”, C.W. Potter, J. of applied microbiology, 2001, 91,572—579.
Parts of the course will include material from the
more advanced material in the book:
“Networks, Crowds, and Markets:
Reasoning About a Highly Connected
World” by David Easley and Jon Kleinberg.
– A preprint version is available online at:
http://www.cs.cornell.edu/home/kleinber/networks-book/

Dueling algorithms

Transcription

Similar documents

Dueling algorithms