Professional Documents
Culture Documents
4.4A-2-1
ITC 11 Kyoto September 1985
4.4A-2-2
ITC 11 Kyoto September 1985
w
Thus, to evaluate dJ/da one needs only to either the single- or multi-choice type; and
identify the Ki~~ 's uV at the operating point. d) network conditions are continuously monitored
UvW
Recalling that these coefficients actually repres- such that its data is available for outputting
ent the slopes of P~. 's with respect to a
W
's, as desired o
one may justifiably1 use the following li~~ar
J
The network used for the simulation results
recursive estimator
reported below, has 3 source nodes, 4 transit
..~ nodes, and 3 sink nodes. The trunk group capaci-
K1J (n-l) 0 (l-y ) + ties and the load pattern are as given in Tables
uVW ~ 2
~ 1 and 2, respectivelyo The average call duration
P .. (n) P .. (n-l)]
[
1J 1J .Y is T = 100 sec and the permitted routing decisions
w _ a W (n-l) 2 are of the single-choice load-sharing type. After
a (n)
UV uV normal operation for a period of 1150 sec, the
(11) failure of the transit exchange corresponding to
together with k = 1 is simulated. This condition is maintained
for 850 sec during which calls COUld. be routed
if the (m+l) st call through transit nodes 2 and 3 only.
succeeds
A number of simulation experiments were then
if the (m+l) st call
carried out using the proposed routing algorithmo
is blocked
In one experiment, the algorithm performance was
(12)
compared to that of the optimal solution detailed
In the above, n is an index for the updating in Section 3 0 For this purpose, several initial
instant, whereas m is a call dependent index. trials were conducted in order to determine suite
Moreover, Y and Y are constants whose inverses able values for the new algorithm's parameters o
l 2
correspond Eo the estimators' depth of memory. This yielded: Y = 003, Y = .0012, a = 016, and
l 2
meanwhile, P denotes the blocking probability of d = 50 sec. The average number of lost calls over
a trunk group belonging tothe route under 100 replications of simulation runs (each having
consideration. 2000 sec of simulation time), was then obtained
and the results are shown in Figo 1. Also shown
Next, to tackle the problem of centraliza-
in the figure are the results of the off-line
tion, Eqo (9) is approximated as follows :
optimal solution for the routing problem, as ind-
For all i,j £ {l, •• o,s} icated with the dotted lines. It is clear from
Fig. 1 that the automata were able to rapidly
dJ dJ
a~. (n+l) a {a~ . (n) -l} i f - - =min--
"learn" the routing strategy that should be foll-
1J 1J da~. V da~. owed under each network condition, and that these
strategies are very close to the optimal ones.
1J 1J
a~j(n) - a{a~j(n)-o} otherwise
In another simulation experiment, the effect
of varying the values of Y , Y , a, and d was
(13) l 2
studied. For this purpose, the following set of
(The above approximation says that we replace the reference parameter values was chosen: Y = 007,
l
f.I Y = .001, a = 0.1, and d = 50 sec o The results
boundary point a .. (n) - --E dJ/da . . (n) by the
-J..J a -J..) o~tained by separately varying each of these par-
vertix nearest to it in the feasible set.) ameters are shown in 'Figs. 2 - 6. Based on these
figures one may draw the followin conclusions :
Equation (13) represents a linear automaton i) Increasing the depth of memory - which is given
of the reward-inaction type with a learning step by l/Y and 1/Y - leads to a reduction in the
l 2
"a" [lOJ. It is exclusively concerned with the responsiveness of the algorithm to random fluc-
routing of calls between nodes i and j. More- tuations, thereby reducing the variance of
over, by making the realistic assumption that losses at steady state;
ij~ . .
K ~ 0 for those nodes wh1ch are phys1cally ii)Increasing the learning step "a" causes a more
d~~~ant from i and j, the automaton needs only rapid convergence in the event of large amplit-
to receive traffic data frome those nodes in its ude perturbations; however, this is accompanied
immediate vicinity. (See Eqo(lO». Consequently, by an increase in the variance of losses at
decentralization of routing actions can be achieved. steady state;
iii) Decreasing the updating period" d" has an infl-
5. PERFORMANCE EVALUATION VIA COMPUTER uence similar to that of increasing "a".
SIMULATION
It is therefore clear that when choosing the
In order to evaluate the performance of the parameters of the proposed routing algorithm,
a trade-off has to be made between the speed of
proposed routing algorithm" a special interactive
convergence and the steady state performance. In
computer program has been developed in PASCAL [llJ.
It has the 'following features : a) the call inter- addition, the choice of the parameter "d" should
take into account the permissible rate of data
arrival times are exponentially distributed in
accordance with the offered traffic matrix {A .. }; transmission and processing, since this sets
b) the call durations are exponentially distrt~uted a lower limit to "d"o
with an average value T which is the same for
all routes; c) routing strategy is updated period-
icallyevery "d" seconds according to Eq. (13),
with a facility for having routing actions of
4.4A-2-3
ITC 11 Kyoto September 1985
In this paper we have presented a new routing [lJ P. J. Burke, "Automatic overload controls in
algorithm which offers a number of significant a circuit-switched conununications network",
advantages over previously suggested algorithms. in Proc. Nato Electron. Conf., pp. 667-672,
On the one hand, it overcomes the prohibitive December 1968.
computational burden and the inescapable central-
ization requirements, associated with the optimal [2 J L. Ao Gimpelson, "Network management : design
solution of the routing problem using gradient and control of conununications networks",
techniqueso On the other hand, it avoids the need Electrical Conununications, vol o 49,No o 1,
for processing traffic data at the impractically pp. 4-22, Jan. 1974.
high "call-by-call" rate demanded by standard
learning automata techniques [3], [la]. This is [3J K. S. Narendra, E. A. Wright, and Lo G. Mason,
accomplished while maintaining a near optimal "Application of learning automata to telephone
performance that cannot be guaranteed using stand- traffic routing and control", IEEE Trans. on
ard learning automata techniques, especially under Syso, Man, and Cyb., vol SMC-7, No. 11, pp.
heavy traffic conditions o 785-792, November 1977.
4.4A-2-4
ITC 11 Kyoto September 1985
50
45
40
35
III
0'1
C 30
cO
r-i
H
~ 25
u
~ 20
~
E-t 15
E-t
Ul
o
H 10
o 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
TIME ( 100 SEC.)
Fig. 1 Behaviour of the new routing algorithm including effect of breakdown
50
45
40
35
••• •
~ 30
c
cO
g 25
r-i
u
~ 20
~E-t 15
E-t
Ul
o
H 10
O~~~~--~--~--~~--~--~--~9--~1:0--~11~~1~2~~1;3~1~4~~155~1~6--tl7~]1~8~~19;--;20
o 1 2 3 4 5 6 7 8 TIME ( 100 Sec.)
on the performance of the new routing algorithm
Fig. 2 The effect of parameter Yl
4.4A-2-5
ITC 11 Kyoto September 1985
50
45
40
35
.001
III •• • • Y2
g' 30 .00085
1\1
-O-~~-<>- Y2
r-i
H
~ 25
u
H 20
rr..
~8 15
8
Ul
0
...:l 10
o 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
TIME ( 100 Sec.)
Fig. 3 The effect of parameter Y2 on the performance of the new routing algorithm
50
45
40
35
• • •• a .1
-<>-o--o-~ a .08
U
H
rr.. 20
~8
15
8
Ul
o
...:l
10
o 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 . 17 18 19 20
TIME ( 100 Sec.)
Fig. 4 The effect of decreasing a on the new routing algori tm behaviour
4.4A-2-6
ITC 11 Kyoto September 1985
50
45
40
\
\
35
\
III
\
0'1
• • •• a .1
~
~ 30 I
rl a
H -0- -0- -0- -0- .25 "'t:J
~
25
u
H
rz..
20
~
E-<
E-< 15
Ul
0
H
10
0 1 2 3 4 5 6 7 8 9 10 11 12 13
15 16 17 18 19 20 14
Fig. 5 TIME ( 100 Sec.)
The effect of increasing a on the new routing algorithm behaviour
50
45 ~,
'0. .......
40 ""'0.
\ ,
\
35
•• •• d 50 sec.
d 100 sec.
U
1
I
H
rz.. 20
~
E-< I
E-< 15 I
Ul
o
H I
10
o 1 2 3 4 5 6 7 8 . 9 10 11 12 13 14 15 16 17 18 19 20
~IME ( 100 Sec.)
Fig. 6 The effect of slow updating on the new routing algorithm behaviour
4.4A·2~7