ga-fca
TRANSCRIPT
-
8/2/2019 GA-FCA
1/35
Channel Assignment
in Cellular Networks
Ivan Stojmenovic
www.site.uottawa.ca/~ivan
-
8/2/2019 GA-FCA
2/35
Overview
Fixed channel assignment
Multicoloringco-channel interference
General problem statement
Genetic algorithms
Results and details Fixed/dynamic channel and power assignment
-
8/2/2019 GA-FCA
3/35
Cell structure Implements space division multiplex: base station
covers a certain transmission area (cell) Mobile users communicate only via the base station
Advantages of cell structures:
higher capacity, higher number of usersless transmission power needed
more robust, decentralized
base station deals with interference locally Cell sizes from some 100 m in cities to, e.g., 35 km
on the country side (GSM) - even more for higher
frequencies
-
8/2/2019 GA-FCA
4/35
Cellular architecture
One low power transmitter per cel
Frequency reuselimited spectrum
Cell splitting to increase capacityA B
Reuse distance:minimum distance betweentwo cells using same channel for satisfactory
signal to noise ratio
Measured in # of cells in between
-
8/2/2019 GA-FCA
5/35
Problems
Propagation path loss for signal power: quadratic or higher indistance
fixed network needed for the base stations
handover (changing from one cell to another) necessary
interference with other cells:
Co-channel interference:
Transmission on same frequency
Adjacent channel interference:
Transmission on close frequencies
-
8/2/2019 GA-FCA
6/35
Reuse pattern for reuse distance 2?
One frequency can be (re)used in all cells of the same color
Minimize number of frequencies=colors
-
8/2/2019 GA-FCA
7/35
Reuse distance 2reuse pattern
One frequency can be (re)used in all cells of the same color
-
8/2/2019 GA-FCA
8/35
Reuse pattern for reuse distance 3?
-
8/2/2019 GA-FCA
9/35
Reuse distance 3reuse pattern
-
8/2/2019 GA-FCA
10/35
Frequency planning I
Frequency reuse only with a certain
distance between the base stations Standard model using 7 frequencies:
Note pattern for repeating the same color:
one north, two east-north
f4
f5
f1f3
f2
f6
f7
f3f2
f4f5
f1
-
8/2/2019 GA-FCA
11/35
Fixed and Dynamic assignment
Fixed frequency assignment: permanent certain frequencies are assigned to a certain cell
problem: different traffic load in different cells
Dynamic frequency assignment: temporary base station chooses frequencies depending on the
frequencies already used in neighbor cells
more capacity in cells with more traffic
assignment can also be based on interferencemeasurements
-
8/2/2019 GA-FCA
12/35
3 cell cluster
with 3 sector antennas
f1f1 f1
f2
f3
f2
f3
f2
f3h1
h2
h3g1
g2
g3
h1h2
h3g1
g2
g3g1
g2
g3
-
8/2/2019 GA-FCA
13/35
Cell breathing
CDM systems: cell size depends on current load
Additional traffic appears as noise to other users
If the noise level is too high users drop out of
cells
-
8/2/2019 GA-FCA
14/35
Multicoloring
Weight w(v) of cell v = # of requested frequencies
Reuse distance r
Minimize # channels used: NP hard problem
Multi-coloring = multi-frequencing
Channel= Frequency= Color
Hybrid CA = combination fixed/dyn. frequencies
Graph representation: weighted nodes, two nodesconnected by edge iff their distance is < r
same colors cannot be assigned to edge endpoints
-
8/2/2019 GA-FCA
15/35
Hexagon graphs: reuse distance 2
What is the graph for reuse distance 3?
-
8/2/2019 GA-FCA
16/35
Lower bounds for hexagonal graphs
D= Maximum total weight on any clique
Lower bound on number of channels: DD/3
D/2 D/6
D/2
D/2 D/2
D/2
D/2
D/2D/2D/2
D/2000
Odd cycle bound: induced 9-cycle, each weight D/2
Channels needed in this cycle: 9D/2
Each channels can be used at most 4 times.
Needs 9/8D
channels
-
8/2/2019 GA-FCA
17/35
Fixed allocationsreuse distance 2D= maximum number of channels in a node or 3-cycle
Red : 1, 4, 7, 10, Green: 2, 5, 8, 11, Blue: 3, 6, 9, 12,
Total # channels: 3D Performance ratio: 3
Janssen, Kilakos, Marcotte 95: D/2 red, blue and green each
D/2
D/2
D/2
Each node takes as many channels as neededfrom its own set
If necessary, RED borrow from GREEN
BLUE borrow from RED
GREEN borrow from BLUE
If a node has D/2+x channels, no
neighbor has more than D/2-x channels
3D/2 channels used, performance ratio: 3/2
-
8/2/2019 GA-FCA
18/35
-
8/2/2019 GA-FCA
19/35
Feder-Shende algorithm-reuse dist. 3
Base color underlying graph with 7 colors
Assign L channels to each color class
Every node takes as many channels as it needs from
its base color set Heavy node (>L colors) borrows any unused
channels from its neighbors
L=D/3
algorithm with performance ratio 7/3 Reuse distance r perform. ratio 18r2/(3r2+20)
2: 2.25, 3: 3.44, 4: 4.23, 5: 4.73 (Narayanan)
k-colorable graph perf. ratio k/2(Janssen-Kilakos 95)
-
8/2/2019 GA-FCA
20/35
Adjacent channel interference
Receiver filter
f1 f3f2interference
Co-site constraint: channels in the same cell must be
c0 apart
Adjacent-site constraint: channels assigned toneighboring cells must be c1 apart
Inter-site constraint: channels assigned to cells that are
r cells apart must be crapart
-
8/2/2019 GA-FCA
21/35
Lower bounds: co-site and adjacent-site
Gamst 86
c0max {w(u), w(v), w(x)}
c1 max{vCw(v) | C is a clique}
max {c0 w(u), (c0c1)w(u)+c1vC,vu w(v) | C is a
clique containing u} when c02c1
u
v x
c0c1c0
-
8/2/2019 GA-FCA
22/35
3-colorable graphsDistance between channels = max(c0/3, c1)
Borrowing impossible
Distance between channels = max(c0/2, c1)
Borrowing possible
Borrowed channels = change colordynamic CA=online distributed CA
Channels with ongoing calls can(not) be borrowed = (non)recoloring
k-local algorithm: node changes channels based on weights within kcells
-
8/2/2019 GA-FCA
23/35
Desirable qualities of CA algorithms
Minimize connection set-up time
Conserve energy at mobile host
Adapt to changing load distribution
Fault tolerance Scalability
Low computation and communication overhead
Minimize handoffs Maximize number of calls that can be accepted
concurrently
R h bl l l l
-
8/2/2019 GA-FCA
24/35
Research problem: several power levels
at mobile hosts
If mobile phone is near base station, it may switchto lower power level
Interference from other hosts increases
Interference of that host to other node decreases Are there benefits of using two power levels?
Fixed or dynamic channel and power assignment
and multicoloring: simplest cases Fixed or dynamic channel and power assignment
with co-site, adjacent-site and inter-site constraints:
Genetic algorithms, simulated annealing,
-
8/2/2019 GA-FCA
25/35
Genetic algorithms
Rechenberg 1960, Holland 1975
Part of evolutionary computing in AI Solution to a problem is evolved (Darwins theory)
Represent solutions as a chromosomes = search space
Generate initial population of solutions(chromosomes) at random or from other method
REPEAT
Evaluate the fitness f(x) of each chromosomex Perform crossover, mutation and generate new
population, usingf(x) in selecting probabilities
UNTIL satisfactory solution found or timeout
-
8/2/2019 GA-FCA
26/35
Fixed channel assignment problem
INPUT: n = number of cells
Compatibility matrix C, C[i,j]= minimal channelseparation between cells i and j, 1i,jn
d[i] = number of channels demanded by cell i
OUTPUT: S[i,k]= channel # of k-th call of cell i, 1kd[i] CONSTRAINTS: |S[i,k]-S[j,L]|C[i,j],1kd[i], 1Ld[j], (i,k)(j,L)
GOAL: minimize m= max S[i,k] = # channels
reducable to graph coloring problem NP-complete GA solution space: m fixed, F[j,k]=0/1 if channel k
is not assigned/assigned to cellj, 1km, 1jn.
Optimization: Minimize number of interferences and satisfy demand
O bl t ti d
-
8/2/2019 GA-FCA
27/35
Our problem representation and
solution space
Each row F[j,k], 1km, is a combination ofd[j] outofm elements (# of 1s is = d[j])
Cost function to minimize: C(F)= A+B
A= total number of co-site constraint violations B= total number of adjacent and inter-site violations
= parameter; C(F)=0 for optimal solution
Initial population: generate restricted combinations:
generate random combination ofd[j] Xs and m-
(c0+1)d[j] 0s; replace each X by 100..0 (c0 0s);
shift circularly by random number in [0,c0]
-
8/2/2019 GA-FCA
28/35
Mutation
Each row=cell is mutated separately
Combinations in bit representation: x 1s out ofm bits
Mutation with equal probability for each bit: choose one out
ofx 1s and one out ofm-x 0s at random, swap: Ngo-Li 98
Mutation with different probability for each bit:b[i]= # of conflicts ofi-th selected channel
with other channels in this and other cells
p[i]=b[i]/(b[1]++b[x])
Repeat for 0s: # of conflicts if that channel turned on Choosing bit with given probability:
Generate at random r, 0 r1, and choose i,
p[1]+p[i-1] r
-
8/2/2019 GA-FCA
29/35
Crossover Regular GA crossover:
1011000110 10011110000101111000 0111000110
Ngo-Li 98:A and B two parents, each row separately,
preserve # of 1s in each row:push 10 and 01 columns in stack if top same;
pop for exchange if top different
1011000110 1001101000
0101111000 0111010110
Problem: # of swaps varies
-
8/2/2019 GA-FCA
30/35
New crossover
t= number of desired swaps in a row
Mark positions in two combinations that differ
let s 10s and s 01s are found
Choose tout ofs 10 at random and
01 Choose tout ofs 01 at random and 10
Example: 1011000110 1001010010
0101111000
0111110110s=4 t=2 $^$ ^^^$$ # **#
# **# offspring
selected columns
-
8/2/2019 GA-FCA
31/35
Crossover needs further study
Problem: independent changes in each row=cell will
destroy good channel assignments of parents
Two good solutions may have nothing in common
Try experiments with mutation only
(may be crossover has even negative impact !?)
Evaluate impact of each column change by cost
function and apply weighted probabilities for
column selections
Best value for tas function ofs? t=s/2? Small t?
-
8/2/2019 GA-FCA
32/35
Combinatorial evolution strategy
Sandalidis, Stavroulakis and Rodriguez-Tellez 98
Generate individuals and evaluate them byf
Select best individualindiv; indiv1=indiv; counter=0; t=0;
REPEAT t=t+1 IF counter=max-countTHENapply increased mutation rate
(destabilize to escape local minimum)
Generate individuals from indiv1 and evaluate them byf
Select best individual indiv2 IF indiv2 better than indiv1 THEN {counter=0; indiv=indiv2} ELSE
{counter=counter+1; indiv1=indiv2}
UNTILtermination
Applied for fixed, dynamic and hybrid CA
-
8/2/2019 GA-FCA
33/35
CES for dynamic channel assignment
n=49 cells, m=49 channels, call arrives at cell k
F[j,i]=0/1 if channel i is not assigned/assigned to
cellj, 1im, 1jn: current channel assignment for ongoing calls
Reassignment of all ongoing calls at cell k(channel foreach call may change) to accommodate new call
V[k,i] = new channel assignment for cell k CES minimizes energy function that includes: interference of new
assignment, reusing channels used in nearby cells, reusing channelsaccording to base coloring scheme, and number of reassignments
Centralized controller
CES for Hybrid CA and for borrowing CA in FCA
-
8/2/2019 GA-FCA
34/35
Simple heuristics for FCA
Borndorfer, Eisenblatter, Grotschel, Martin 98
(4240 total demand, m=75 channels, Germany)
DSATUR: key[i]= # acceptable channels remained in cell i,cost[i,j]= total interference in cell i if channel j is selected
Initialize key[i]= m; cost[i,j]=0; i,j WHILE cells with unsatisfied demand exist DO {
Extract cell iwith unsatisfied demand and minimumkey[i];
Let j be available channel which minimizes cost[i,j];
Update cost[x,y] x,y by adding interference (i,j)
Update key[x] x, reduce demand at cell i }
-
8/2/2019 GA-FCA
35/35
Hill climbing heuristic for FCA Borndorfer, Eisenblatter, Grotschel, Martin 98
Two channel assignments are neighbors if one can be obtained fromthe other by replacing one channel by another in one of cells.
PASS procedure for assignmentA={(cell,channel)}:
Sort all (i,j)A by their interference in decreasing order
FOR each (i,j)A in the order DO
Replace (i,j) by (i,j)if later has same or lower interference
Hill climbing for FCA: initializeA; A=A REPEAT
A=A; A= PASS(A)
UNTIL A A i t f (A)i t f (A)