Consensus Experiments

Source code is at github: git@github.com:worden-lee/consensus-simulation.git and git@github.com:worden-lee/adap-dyn.git.

6/14/2011

I did a very preliminary "experiment": a "consensus process" with one person is able to find a local fitness peak. Actually, this is an important baseline: I'll probably want to know how often it finds an acceptable solution, how good it is, etc.

6/15/2011

What are some initial experiments to do?

1.

• $n$ people working independently vs.
• $n$ people each generating a proposal and then voting vs.
• $n$ people searching together in some simple way:

Compare mean fitness of outcome, how satisfied each person is.

2.

• replicate Page's result?
• generate something similar but different to Page's - usefulness of people with somewhat different valuations, as opposed to the same valuation but different search heuristics

Revision 1e0b42d of consensus-simulation does the following:

• initial proposal is 0
• repeat:
  • each individual does steepest-ascent search to a local peak on their landscape, offers that as a next proposal.
  • each individual evaluates each next proposal, blocks it if they consider it worse than the existing proposal
  • if any new proposals have no blocks, the first of them is accepted
• until no more progress is made. note this is a deterministic process.

Preliminary runs suggest that for completely uncorrelated landscapes (with themselves and with each other) larger group size implies no progress, due to too many blocks.

Higher-dimensional landscapes may support more progress with somewhat larger groups than smaller landscapes.

This model protocol is somewhat unforgiving: one person's best offer may be crappy for others, while a nearby point nearly as good may be much better received, but we don't give it a chance. We might have much better success if we put those intermediates on the table as well.