1. Motivation Scenario Algorithm Evaluation
A Cognitive-Inspired Model for Self-Organizing
Networks
ASENSIS 2012
Daniel Borkmann0 Andrea Guazzini12 Emanuele Massaro3
Stefan Rudolph4
0
Communication Systems Group, ETH Zurich, Switzerland
1
Institute for Informatics and Telematics, National Research Council, Pisa, Italy
2
Department of Psychology, University of Florence, Italy
3
Department of Informatics and Systems, University of Florence, Italy
4
Organic Computing Group, University of Augsburg, Germany
10th September, 2012
Borkmann, Guazzini, Massaro, Rudolph Self-Organizing Networks 1 / 19
3. Motivation Scenario Algorithm Evaluation
Motivation
Large Scale Networks emerge
Internet
Pervasive Computing
Often used: Overlay networks
Problems of overlay networks
Structured: Hard without global information
Unstructured: No optimization of network structure
Idea
Self-optimization of an overlay network
Through a cognitive-inspired model
Borkmann, Guazzini, Massaro, Rudolph Self-Organizing Networks 3 / 19
4. Motivation Scenario Algorithm Evaluation
Scenario
Connected network of n nodes
Static, nodes don’t disappear or appear
Each holdes one item (e.g. a service or data)
Each wants to retrieve items with respect to its energy
Each has a limited number of links from 1 . . . m
Each node can change its links
Optimization problems: change links in order to
Retrieve all items with the minimum number of hops
Maximize the number of items with a fixed amount of hops
Borkmann, Guazzini, Massaro, Rudolph Self-Organizing Networks 4 / 19
5. Motivation Scenario Algorithm Evaluation
Cognitive-Inspired Hub Detection
Diffusion and Competitive Interaction
At start
A is the adjacency matrix
Every node i has a state vector Si (short term memory)
(k )
Si is the probability that node i belongs to community k
Every node belongs to its own community
Update of the state vectors
1
S (t + 2 ) = mSik (t ) + (1 − m) ∑j Aij Sjk (t )
α 1
Sik (t + 2 )
S (t + 1) =
∑j Sij (t + 1 )
α
2
Borkmann, Guazzini, Massaro, Rudolph Self-Organizing Networks 5 / 19
6. Motivation Scenario Algorithm Evaluation
Cognitive-Inspired Hub Detection
Diffusion and Competitive Interaction
Entropy
Ei = − ∑(Sj · log (Sj ))
Plateaus show sub-clusters
When curvature changes sign, save information in temporary
memory box
Shannon entropy of information
6.00
5.00
4.00
Entropy
3.00
2.00
1.00
0.00
0 5 10 15 20 25 30 35 40
Time
Borkmann, Guazzini, Massaro, Rudolph Self-Organizing Networks 6 / 19
7. Motivation Scenario Algorithm Evaluation
Cognitive-Inspired Hub Detection
Cognitive Dissonance
Cognitive concept found by social psychologists
Reduces conflicting cognitions
Creates consistent belief system
∑k |Sik −Sjk |
Here: Dij := 2
Interesting for adaption of α :
Eit −1 +Dit −1 Eit +Dit
If Ki
− Ki
< ε for more than τ ∗ times
Set αi = 1.5|η (0,σ ) | + 1
Borkmann, Guazzini, Massaro, Rudolph Self-Organizing Networks 7 / 19
8. Motivation Scenario Algorithm Evaluation
Cognitive-Inspired Hub Detection
Long Term Memory
Store potential hubs in the Long Term Memory
Find B 1 time positions by sorting with respect to first derivative
Sort the remaining vectors with respect to the entropy
Find the potential hubs in the state vectors
Use Long Term Buffer of size B 2
The last B 2 sets of size B 1 are stored (bounded rationality)
This creates a (B 1 , B 2 ) matrix
Borkmann, Guazzini, Massaro, Rudolph Self-Organizing Networks 8 / 19
9. Motivation Scenario Algorithm Evaluation
Rewiring
With help of this Long Term Memory, we can can create a “hub
list" for each node
Rewiring steps:
1. Find the weakest X % of the nodes
2. Choose Y % of the nodes at random
3. Each of these nodes closes a connection to a non-hub
4. Each of these nodes opens a new connection to a potential hub
Borkmann, Guazzini, Massaro, Rudolph Self-Organizing Networks 9 / 19
15. Motivation Scenario Algorithm Evaluation
Numerical Simulation
Scenarios
1. Maximization of the reachable items of the nodes
The energy (hops) is limited
Weakest nodes: Minimum number of items
2. Minimization of used energy
All item will be reached in every step
Weakest nodes: Maximum number of energy
Randomized Algorithm
For comparison
Does not use hub list
Borkmann, Guazzini, Massaro, Rudolph Self-Organizing Networks 15 / 19
16. Motivation Scenario Algorithm Evaluation
Numerical Simulation
Parameters
Number of nodes n
Mean connectivity
Mean extra connectivity
Number of unique items I
Number of items to retrieve Imax
Hub detection: m, α
Rewiring
Borkmann, Guazzini, Massaro, Rudolph Self-Organizing Networks 16 / 19
18. Motivation Scenario Algorithm Evaluation
Evaluation
Results for the minimization of energy
Topology Optimization
166.00
164.00
162.00
160.00
Mean energy
158.00
156.00
154.00
152.00
150.00
148.00
0 100 200 300 400 500 600 700 800 900 1000
Round
Rewiring, cognitive approach Rewiring, randomzied approach
Setting: Mean over 50 runs, n = 200, mean_conn= 4, extra_conn= 4, I = 50, Imax = 45, rw_weak= 0.09, rw_rand= 0.03
Borkmann, Guazzini, Massaro, Rudolph Self-Organizing Networks 18 / 19
19. Motivation Scenario Algorithm Evaluation
Conclusion
Contributions
Development of a cognitive model for community detection
Application of information for self-optimization of a network
Comparison with a randomized algorithm
Future Work
(i) Evaluate the algorithm on a wide range of large scale network
topologies
(ii) Localize the decision making of a node when to rewire or not
(iii) Introduce more dynamics into items and nodes
Borkmann, Guazzini, Massaro, Rudolph Self-Organizing Networks 19 / 19