k Is(0) = s ), k = 0, 1, ... ; s ~ S; for every ~.
(2)
Since the time between transitions is independent of both the state of the system and the control policy, and since the state does not change between transitions, inequalities (2) are equivalent to P{
> k Is(0) -- s }
k=0, 1.... ; n=l,
2.... ;
> k Is(0) = s
s ~ S ; for every ~.
}, (3)
Assume that ~r° is a deterministic policy and that the action prescribed by ~r° in state s is denoted by
~-°(s).
P.K. Johri / Minimizing the number of customers in queuing systems
113
Theorem 1.7r o stochastically minimizes the mtmber of customers in the system if it satisfies (3) for n = 1 and if
~_, q'(s'ls, ~ r ° ( s ) ) [ P { N . o ( n ) > k l s ( O ) = s ' } - P {
N.o(n) > kls(O)=s}]
$'ES
<~ E q'(s'ls, a ) [ P { N , o ( n ) > k l s ( O ) = s ' } - P { N , , o ( n ) > k l s ( O ) = s } ]
,
s'~S
aEA,; k=0,1 .... ; n=0,1 .....
(4)
Remark. Inequalities (4), since they are only in terms of ~r°, are simpler to establish than inequalities (3). Proof. It will be shown, by induction on n, that ~r° satisfies (3) for every n. Assume that it does for some n = n' >~ 1, which implies that if only the first n' transitions are of interest then rr ° should be employed. For n = n' + 1, if an action is taken initially according to policy rr °, then the result follows from the induction hypothesis. So consider a policy rr' which initially prescribes an action (or randomized actions) different than the initial action prescribed by ~r° and then is identical to 7r° for the next n' transitions. By conditioning on the state after the first transition and using the Markov property, P{N,~o(n'+ 1) > k l s ( 0 ) = s } = P { N = o ( n ' + 1) > k l s ( 1 ) = s } + [left hand side of (4) with n = n ' ] / A , whereas P ( N ~ . ( n ' + 1) > k Is(0) = s } = P ( N = , ( n ' +
1) > k I s ( l ) = s }
+ [right hand side of (4) with n = n ' ] / . 4 . It follows that ~r° satisfies inequalities (3) for n = n' + 1 and the induction is complete.
Optimal customer selection in an M / M / 1
queue ssJth feedback
For the first model described previously, the state of the system can be denoted by a vector s = (s~, s 2 . . . . . s,,,) where s~ = j if there are j customers of type i in the system, i = 1, 2 . . . . . m; j = 0, 1. . . . . Given a state s, let
C(s)
=(i:si>O},
( + 1 i , s)=(sx ..... si_l, s~ + l, si+ 1..... s,,), Isl
=s~ + s 2 +
--- +s,..
Let ~ = ~ t l "]- /'L2 -[" " ' " q - ~ m ,
~k = ~ k l "{- ~k2"}- " ' " -l-~krn,
and assume, without loss of generality, that PlPl >/P2#2 >/ "" " >/Pm/X.,, and that A=k+~=I. Define ~r- to be the policy which serves any customer with the least index among the customers present, that is, the action prescribed by ~r in state s can be denoted as serving a type ~r-(s) customer where =
i
C(s).
114
P.K. Johri / Minimizing the number of customers in queuing systenu
It is obvious that ~r satisfies ineqs. (3) for n = 1. Ineqs. (4) simplify to
0+,.+,[ P{ g~-(n) > kls(O)= ( - 1 + .
s)} - P{ g~-(,,)> k i s ( o ) = s
}]
< p;.j[ P { No-(,,) > ~ Is(O) = ( - ~, s) } - p (N~-(,,) > ~. s(O) = s }], (5)
s ~ S ; for everyj ~ C ( s ) ; n=O, 1 . . . . ; k=O, 1 . . . . . Define
h.(s; i; k) = pibti[P(N=-(n) >
kls(O)=
(-1.
s)} -
P{ N,~-( . ) > kls(O)
Lemma 1. For every s ~ S, for every i, j ~ C(s) such that i < j , n = 0, 1. . . .
=s}].
and k = O, 1. . . .
h.(s; i; k)<~h.(s; j ; k)~
s)
and
u=~r-(-lj,
s)=rr-(s)<~i.
Case (i): u = i. By conditioning on the state after the first transition, and using the Markov property, h.,(s; i; k) = pd~,lP{N.-(n' ) > k Is(l) = ( - l v , - 1 , , s)}po~o +P{N,,-(n') > kls(1 ) = ( - 1 / , s ) ) ( # - pj%)
+ ~P(N.-(n')>kls(1)=(lr,-li,
s)}X r
r=l
- P { N,; (n') > k ls(1 ) = ( - 1i, s)} P,#i
-P{
N~-(n') > k I s ( l ) = s } (/~ - p,/t,)
-- r~= l
P{N~-(n')>kls(1)=(l~'s)}}~r]
=p,tt,h.._l((-l,, s); v; k)+(tt-pd~i)h.._,(s; i; k) + ~ )~rh.,_l((l,. s ) ; i; k ) . r~l
Similarly,
h,,(s; j; k ) = p d x i h w _ l ( ( - l i , s); j ; k)+(tt-Pitti)h,,_l(s; j; k) m
+ E ?,,h,,_,((a,, s); j; k). r=l
It follows from the inductive hypothesis that the result is true for n = n'.
(6)
P.K. Johri / Minimizing the number of customers in queuing systems
115
Case (ii): u < i. For this case, since v = u, the induction is simple and is omitted. The proof of lemma 1 is complete. It has just been shown that rr satisfies the conditions of Theorem 1 and hence, ~r stochastically minimizes the number of customers in the system at any time t > O. Optimal assignment of sen'ers in an
M/M/m
queue
For the second model described earlier, the state of the system is denoted by (N, s) where N equals the number of customers in the system and s = (s 1, s 2..... sin) is a binary vector such that s i = 1 (0) if server i is busy (idle). Given a state (N, s), let C(s) be the set of indices of busy servers, C'(s) be the set of indices of idle servers, ~(s) = Ei~ c(s)t~;, s i = (sl ..... s~-l, 1, s,+l . . . . . s,.), i, = ( s a. . . . . si_l, O, si+ 1..... s,,). Without loss of generality assume that /q >~/x2>~ - - - >~t,~, and that A =~,+/.tl +/.t2+ - - - + / I r a = 1 . Derman et al. (1980) have shown that the optimal poficy is to always assign a customer to that free server whose service rate is largest. So let ~r^ be the policy which always sends a customer to the ,r^(s)-th idle server, where ~r^(s) = m i n i ~ C'(s), whenever servers are free. It is obvious that ~r^ satisfies ineqs. (3) for n = 1. Since servers are not permitted to be idle when a customer is waiting, options exist only when a new customer arrives and two or more servers are free. To distinguish such states denote them as (1, N, s). If this new customer is assigned to server j then the new state is ( N + 1, sJ). For the purpose of establishing ineqs. (4), which are in terms of ~" only, the new state can be written as ( N + 1, s ) where s ~= s *i'). Ineqs. (4) can be written as
X[P{N;(n) > k I,(0)= (1, N + 1, s^)} - P { N ; (,,)> k Is(0)= (N+ 1, ,^)}] + Z rEC(s') s J ) } - P ( N . - ( n ) > k l s ( O ) = ( N + t . s^)}] ~.[P(N~r'(n)>kls(O)=(N,r(sJ))}-P{g~r'(tl)>kls(O)=(g+l,s')}],
+
E reC(sJ) j~C'(s);(1,
kls(0)=(1.
N,s)~S;
N + 1.
n = O , 1. . . . ; k = O , 1.....
which simplify to #.i,) [ P { N , , - ( n ) > k ls(0 ) = ( N , s)} - P ( N , , - ( n ) > k Is(0) = ( N + 1, s ' ) } ] Z I~.P(N~'(n) > kls(O) = ( N , ~C(s)
+XP{N,r(n ) > k l s ( 0 ) = ( N + 2, ( s ' ) " ) } +
rEC(s)
j~C'(s);(1,
N,s)~S;
n=0,1 .... ;k=O,
1.....
r(s^))}
116
P.K. Johri / Mininzizing the number of customers in queuing systems
Lemma 2. For everyj ~ C'(s), (1, N, s) ~ S, n = 0, 1 . . . .
and k = 0, 1 . . . . .
P ( N ~ - ( n ) > k Is(0) = ( N + 1, s ' ) } < P ( N ~ - ( n ) > k l s ( 0 ) = ( N + 1, s J)} and
P{ N~-(n) > k Is(0) = (N, s)} -P{
N . , - ( n ) > k I s ( 0 ) = ( N + 1, s ' ) } < 0.
The proof, by induction on n, is similar to the proof o f Lemma 1 and is omitted. Since #~(s)>~#j, from Lemma 2 it follows that ~r satisfies ineqs. (4) and hence it stochastically minimizes the number of customers in the system at any time t > 0.
References Crabill, T.B. (1974), "Optimal control of a maintenance system with variable service rates", Operations Research 22, 736-745. Derman, C., Lieberman, G., and Ross, S.M. (1980), "On the optimal assignment of servers and a repairman", Journal of Applied Probability 17, 577-581. Johri, P.K., and Katehakis, M.N., "Scheduling service in tandem queues attended by a single server", OR Spektrum (submitted). Larsen, R.L., and Agrawala, A.K. (1983), "Control of a heterogeneous two-server exponential queuing system", IEEE Transaction on Software Engineering 9 (4), 522-526. Lippman, S.A. (1975), "Applying a new device in the optimization of expor~ential queuing systems", Operations Research 23, 687-710. Nash, P. and Weber, R.R. (1982), "Dominant strategies in stochastic allocation and scheduling problems", in: M.A.H. Dempster, J.K. Lenstra and A.G.H. Rinnooy-Kan (eds.), Deterministic and Stochastic Scheduling~ Reidel, Dordrec:ht, 385-398. Schrage, L. (1968), "A proof of the optimality of the shortest remaining service time discipline", Operations Research 16, 687-690. Varaiya, P. Walrand, J., and Buyukkoc, C. (1983), "Extension to the multi-armed bandit problem", in: Proceedings of the 22rid IEEE Conference on Decision and Control, San Antonio, TX; 1179-1180. Weber, R.R. (1978), "On the optimal assignment of customers to parallel servers", Journal of Applied Probability 15, 406-413. Winston, W.L. (1977), "Assignment of customers to ser~'ers in a heterogeneous queuing system with switching", Operations Research 25, 469-483. Winston, W.L. (1979), "Optimality of the shortest line discipline", Journal of Applied Probability 15, 181-189.