UNIVERSITY OF JYVÄSKYLÄ
Topology Management in Unstructured
P2P Networks Using Neural
Networks
Presentation for IEEE Congress on Evolutionary Computing
27.9.2007
Annemari Auvinen, research student
Department of Mathematical Information Technology
University of Jyväskylä, Finland
http://www.mit.jyu.fi/
cheesefactory
With co

authors Teemu Keltanen and Mikko Vapa
UNIVERSITY OF JYVÄSKYLÄ
Topology Management Algorithms
•
Topology management algorithms affect the
logical topology by making network more
scalable and effective for resource discovery
•
Use local information the nodes are collecting
about their neighbors
–
Interest based clustering
–
Technical
characteristics
of
the
peers
UNIVERSITY OF JYVÄSKYLÄ
NeuroTopology
•
Uses evolutionary neural networks to form
efficient P2P topologies for resource queries
•
We determine the characteristics that the
neural network should take into account
–
These characteristics are given to the neural
network as inputs and can be e.g. bandwidth or
information about the previous resource queries
•
As a result is obtained dynamic P2P network,
where the topology takes shape in interaction
with the resource discovery algorithm
UNIVERSITY OF JYVÄSKYLÄ
NeuroTopology
•
Algorithm is executed in every peer after a
predefined amount of resource queries
•
Algorithm goes through all neighbor
candidates
•
To establish a connection mutual agreement
from both nodes is needed
UNIVERSITY OF JYVÄSKYLÄ
NeuroTopology
Keep neighbor?
New neighbor?
Neighbor Node
Neighbor’s neighbor
P2P
Node
UNIVERSITY OF JYVÄSKYLÄ
Structure of NeuroTopology
UNIVERSITY OF JYVÄSKYLÄ
Training Program
•
Neural network weights define how neural network
behaves so they must be adjusted to right values
•
This is done using iterative optimization process
based on evolution and Gaussian mutation
Define the
P2P network
conditions
Define the fitness
requirements
for the algorithm
Create candidate
algorithms
randomly
Select the best
ones for next
generation
Breed a new
population
Finally select the
best algorithm for
these conditions
Iterate
thousands
of
generations
UNIVERSITY OF JYVÄSKYLÄ
Neural Network Optimization
•
Evolutionary computing for optimizing the weights
•
Fitness of the used neural network is defined based
on the amount of traffic in the P2P network.
–
Algorithm should locate half of the available resources for
each query
–
Algorithm should use as minimal number of packets and
create as minimum
number of new connections
as possible
•
Mutation is based on the Gaussian random variation
and uses the weighted mutation parameter to
improve the adaptability of the evolutionary search
•
Random variation function was introduced by Fogel
and Chellapilla[1]
UNIVERSITY OF JYVÄSKYLÄ
Simulation Environment
•
P2P network with 100 peers
•
Resources power

law distributed
•
Breadth

first search (BFS), highest degree
search (HDS) and random walker (RW) were
used as resource discovery algorithms
•
The test case was divided to:
–
Training environment
–
Generalization environment
UNIVERSITY OF JYVÄSKYLÄ
Simulation Environment
•
In the training set each generation is started
with a grid topology P2P network and follows
the algorithm:
1.
Do 20 times
1.
10 random peers execute resource queries
2.
Execute NeuroTopology algorithm in every peer using
information from resource queries
2.
Execute 10 resource queries in the P2P network
3.
Calculate the fitness for the neural network using
information from step 2
UNIVERSITY OF JYVÄSKYLÄ
Simulation Environment
•
Training of the neural networks was done
using the HDS algorithm and the amount of
generations was 5000
•
Generalization set was the same as the
training set, except that resource queries
were executed by every peer in the P2P
network
UNIVERSITY OF JYVÄSKYLÄ
Fitness in training environment
UNIVERSITY OF JYVÄSKYLÄ
Fitness in generalization environment
UNIVERSITY OF JYVÄSKYLÄ
Resource query packets and replies in
generalization environment
UNIVERSITY OF JYVÄSKYLÄ
Topology packets and changes in
generalization environment
UNIVERSITY OF JYVÄSKYLÄ
Failed queries in generalization
environment
UNIVERSITY OF JYVÄSKYLÄ
Simulation Results
•
Tested in grid topology, power

law topology and a
random graph topology with 3 resource discovery
algorithms and with and without NeuroTopology
UNIVERSITY OF JYVÄSKYLÄ
Convergence
•
Changing the inefficient grid topology on the
early rounds and limiting the changes when
the efficient topology has been reached
UNIVERSITY OF JYVÄSKYLÄ
References
[1] K. Chellapilla and D. Fogel. Evolving neural
networks to play checkers without relying on
expert knowledge. IEEE Trans. on Neural
Networks, 10 (6), pp. 1382

1391, 1999.
Comments 0
Log in to post a comment