I am not a routing guru, nor am I a good Java programmer, but I had an
idea I wanted to run by everyone...

Perhaps it would be beneficial instead of just routing a request to the
"best" node to instead score the nodes and sample from a weighted
distribution of scores.  The proposed next-generating routing has the
ability to score based on routing time and successful completions - Why
not sample the 5 best scores then compute Boltzmann transition
probabilities (example below)?  This would ensure most routing requests
go to the best node but there's a finite probability of selecting the
second, third, 4th or 5th best... this may allow the node to discover
new routing paths??  Or if the preferred path is slow maybe it would be
able to find an alternate path?

For example, node 1 gets a score of 15, node 2 gets a score of 5 (lower
scores are better).  We arbitrarily pick a "temperature" of 10K [what
this does will become obvious in a little bit], and similar to a process
found in statistical mechanics calculate a total partition function:

q=sum (exp(-score/T))   [sum over all nodes considered]

In this case:

q=exp(-15/10)+exp(-5/10)=0.82966

And the probability of choosing a particular node is the Boltzmann
factor over the partition function:

p1=exp(-15/10)/q= 26.89%
p2=exp(-5/10)/q=  73.10%

The "temperature" is a user adjusted parameter which controls the width
of the distribution.  At higher "temperatures" the probability of
picking a sub-optimal node increases (at 20K p1=37.754% p2=62.22%) --
the distribution widens.  At lower temperatures the system reduces to
the current routing scheme.

Just a thought.  It's loosely based on a technique called simulated
annealing which uses a "temperature" quenching schedule to optimize
either a structure or objective function, first introduced by Metropolis
around 1951 to determine liquid structures.

-- J

_______________________________________________
Tech mailing list
[EMAIL PROTECTED]
http://hawk.freenetproject.org:8080/cgi-bin/mailman/listinfo/tech

Reply via email to