Experiments in Behavioral Network Science: Brief Coloring and Consensus Postmortem Networked Life

Experiments in Behavioral Network Science:
Brief Coloring and Consensus Postmortem
(Revised and Updated 4/2/07)
Networked Life
CSE 112
Spring 2007
Michael Kearns & Stephen Judd
Summary of Events
Held 18 coloring and 18 consensus experiments
Each one had a different network structure (details withheld for now)
All 18 coloring experiments globally solved
17/18 consensus experiments globally solved
average duration ~ 62 seconds
So everyone made $70 or $72
average duration ~ 35 seconds
not bad for less than a couple of hours of “work”
no promises for next round…
Recall (worst-case) status of problems for centralized computation
Seems it is easier to get people to disagree than to agree
Let’s look at the Gallery of Consensus Art
[1] [2] [3] [4] [5]
Network Formation Model
(added 4/2/07)
Single parameter p (a probability)
p=0: a chain of 6 cliques of size 6 each (see figure)
p>0: each intra-clique edge is “rewired” with probability p:
first flip p-biased coin to decide whether to rewire
if rewiring, choose one of the endpoints to “keep” the edge
then choose new random destination vertex from entire graph
Values of p used: 0, 0.1, 0.2, 0.4, 0.6 1.0
Three trials for each value of p; different random network for each trial
S, I attach a revised figure showing how a number of
macroscopic graph quantities vary as a function of p.
In order to plot them all on a common scale and to
compare their rates of change w.r.t. p, for each
quantity I have normalized it so that the minimum
value is always 0 (i.e. subtracted off the min value)
and the max is always 1 (i.e. divided by the max
value, after subtracting off the min). So the y-values
are always between 0 and 1 and thus you can't infer
absolute values, only relative.
The quantities are:
1. average-case diameter (decreases with p)
2. clustering coeff (decreases with p)
3. standard deviation of the degree
distribution (increases with p)
4. maximum degree (increases with p)
5. minimum degree (decreases with p)
6. ratio of the stationary random-walk
probabilities of the most visited
and least visited states (by standard
theory, this is just the ratio of
max to min degrees; increases with p)
Regarding 6., note that we can think of the
stationary probability of a vertex as a measure of
its "centrality", so 6. is a measure of how much
spread there is between the most and least central
As observed before, both 1. and 2. fall off
pretty rapidly with p, which argues for sampling
more finely at small p. On the other hand, 3-6 have
a much wider "range of response".
Sticking with just the consensus games for now, the attached
consplot.jpeg shows a single curve for each consensus game. The curve
shows what fraction of players at any given time are playing the color
that ended up being the "winning" color. You can see that many of the
games have extremely rapid adoption of the winning color, while others
involve quite a long time and lots of "wandering".
S., I am attaching a revision/revisitation of a plot I sent you in the very first results following the experiments. It shows the
average experiment duration (so exactly one of these is at 180 secs, all others completed before) vs. p for both coloring and
consensus on the same plots…
…Digging a little deeper, I compared the completion times of coloring experiments for p = 0, 0.1, 0.2 to those for p = 0.4, 0.6,
1.0. These two sets of completion times pass
a two-sample, unpaired, unequal variance t-test for different means at the P = 0.046 level of statistical significance, beating the
magic standard of 0.05 (i.e. it
is a reportable result by scientific conventions). So there are significant structural effects on coloring performance.
By the way, note that the "flattening" of both curves beyond p=0.4 nicely justifies our spacing of p-values, which recall we did
based on most of the "action" in various structural measures (diameter, clustering coeff, etc.) was at smaller values of p. Nice to
see that thesame seems to hold true of the collective performance.