Transcript Chapter 12: Differential Games, Distributed System, and Impulse Control More than one decision maker, each having separate objective functions which each is trying.

Chapter 12: Differential Games, Distributed
System, and Impulse Control
More than one decision maker, each having separate
objective functions which each is trying to maximize,
subject to a set of differential equations.
The theory of differential games,
Distributed parameter systems,
Impulse control,
Allow to make discrete changes in the state variables
at selected instants of time in an optimal fashion.
12.1 Differential Games
Different types of solutions such as minimax, Nash,
Pareto-optimal, along with possibilities of cooperation
and bargaining.
12.1.1 Two Person Zero-Sum Differential Games
which player 1 wants to maximize and player 2 wants
to minimize.
is the minimax solution.
The necessary conditions for u* and v* ,
is a saddle point of the Hamiltonian function H.
when U=V=E1,
12.1.2 Nonzero-Sum Differential Games
N players,
represent the control
variable for the ith player,
denote the objective function which the ith player
wants to maximize.
A Nash solution:
Open-Loop Nash Solution
for all
Closed-Loop Nash Solution
we must recognize the dependence of the other
player’s actions on the state variable x. Therefore,
Interpretation to the adjoint variable i . Any
perturbation x in the state vector causes them to
revise their controls by the amount
12.1.3 An Application to the Common-Property
Fishery Resource
Let
denote the turnpike (or optimal biomass) level
given by (10.12).
As shown in Exercise 10.2,
We also assume that
which means that producer 1 is more efficient than
producer 2, i.e., producer 1 can make a positive profit
at any level in the interval
, while producer 2
loses money in the same interval, except at
where
he breaks even. For
both producers make
positive profits.
As far as producer 1 is concerned, he wants to attain
his turnpike level , if
If
and if
then from (12.26) producer 2 will fish at his maximum
rate until the fish stock is driven to
At this level it is
optimal for producer 1 to fish at a rate which maintains
the fish stock at level
in order to keep producer 2
from fishing.
The Nash solution;
The direct verification involves defining a modified
growth function:
And using the Green’s theorem results of Section
10.1.2. Since
by assumption, we have
for
From (10.12) with g replaced by g1, it can
be shown that the new turnpike level for producer 1 is
which defines the optimal policy (12.27)(12.28) for producer 1. The optimality of (12.26) for
producer 2 follows easily.
Suppose that producer 1 originally has sole
possession of the fishery, but anticipates a rival entry.
Producer 1 will switch from his own optimal sustained
yield
to a more intensive exploitation policy prior to
the anticipated entry.
A Nash competitive solution involving N  2 producers
results in the long-run dissipation of economic rents.
Model for licensing of fishermen let the control variable
vi denote the capital stock of the ith producer and let
the concave function f(vi), with f(0)=0, denote the
fishing mortality function, for i=1,2,…,N. This requires
the replacement of
in the previous model by f(vi).
Application of differential games to fishery
management,
, Haurie, and Kaitala
(1984,1985) and
, Ruusunen,and Kaitala
(1986,1990). Applications to problems in
environmental management, Carraro and Filar (1995).
Applications of marketing in general and optimal
advertising, Bensoussan, Bultez, and Naert(1978),
Deal, Sethi, and Thompson(1979), Deal(1979),
Jrgensen(1982a), Rao(1984,1990), Dockner and
Jrgensen(1986,1992), Chintagunta and
Vilcassim(1992), Chingtagunta and Jain(1994,1995),
and Fruchter(1999). A survey of the literature is done
by Jrgensen(1982a) and a monograph is written by
Erickson(1991).
For applications of differential games to economics
and management science in general, see the book by
Dockner, Jrgensen , Long, Sorger(2000).
12.2 Distributed Parameter Systems
Systems in which the state and control variables are
defined in terms of space as well as time dimensions
are called distributed parameter systems and are
described by a set of partial differential or difference
equations.
In the analogous distributed parameter advertising
model we must obtain the optimal advertising
expenditure at every geographic location of interest at
each instant of time, see Seidman, Sethi, and
Derzko(1987). In section 12.2.2 we will discuss a
cattle-ranching model of Derzko, Sethi and
Thompson(1980), in which the spatial dimension
measures the age of a cow.
Let y denote a one dimensional spatial vector, let t
denote time, and let x(t,y) be a one dimensional state
variable, Let u(t,y) denote a control variable, and let
the state equation be
For t[0,T ] and y [0,h ]. We denote the region [0,T ]x
[0,h] by D, and we let its boundary D be split into two
parts
and
as shown in Figure 12.1. The initial
conditions will be stated on the part of the boundary
D as
x0(y) gives the starting distribution of x with respect to
the spatial coordinate y. The function v(t) in (12.33) is
an exogenous breeding function at time t of x when
y=0. In the cattle ranching example in Section 12.2.2,
v(t) measures the number of newly born calves at time
t.
Let F(t,y,x,u) denote the profit rate when x(t,y)=x,
u(t,y)=u at a point (t,y) in D. Let Q(t) be the value of
one unit of x(t,h) at time t and let S(y) be the value of
one unit of x(T,y) at time T.
Figure 12.1: Region D with Boundaries
and
12.2.1 The Distributed Parameter Maximum Principle
where xt=x/t and xy= x/y.The boundary conditions
on  are stated for the
part of the boundary of D.
which gives the consistency requirement in the sense
that the price and the salvage value of a unit x(T,h)
must agree.
We let u*(t,y) denote the optimal control function. Then
the discounted parameter maximum principle requires
that
For all (t,y) D and all u  .
These general forms allow for the function F in (12.2)
to contain arguments such as x/ y, 2x/ y2,etc. It is
also possible to consider controls on the boundary. In
this case v(t) in (12.33) will become a control variable.
12.2.2 The Cattle Ranching Problem
Let t denote time and y denote the age of an animal.
Let x(t,y) denote the number of cattle of age y on the
ranch at time t. Let h be the age at maturity at which
the cattle are slaughtered. Thus, the set [0,h] is the set
of all possible ages of the cattle. Let u(t,y) be the rate
at which y-aged cattle are bought at time t, where we
agree that a negative value of u denotes a sale.
Subtracting x(t,y) from both sides of (12.42), dividing
by t, and taking the limit as t 0,yields
The boundary and consistency conditions for x are
given in (12.32)-(12.34). Here x0(y) denotes the initial
distribution of cattle at various ages, and v(t) is an
exogenously specified breeding rate.
Let P(t,y) be the purchase or sale price of a y-aged
animal at time t. Let P(t,h)=Q(t) be the slaughter value
at time t and let P(T,y)=S(y) be the salvage value of a
y-aged animal at the horizon time T. The functions Q
and S represent the proceeds of the cattle ranching
business. Let C(y) be the feeding and corralling costs
for a y-aged animal per unit of time. Let
denote
the goal level purchase rate of y-aged cattle at time t.
The deviation penalty cost is given by
where q is a constant.
subject to the boundary and consistency conditions
(12.38)-(12.40).
Figure 12.2: A Partition of Region D
where g is an arbitrary one-variable function and k is a
constant. We will use the boundary conditions to
determine g and k.
We substitute (12.38) into (12.48) and get
This gives
In the region D1 D2.
For region D3, we use the condition (12.39) on (T,y)
in (12.48) to obtain
(12.50) in the region D1 as the beginning game, which
is completely characterized by the initial distribution x0 .
Also the solution (12.49) in region D3 is the ending
game, because in this region the animals do not
mature, but must be sold at whatever their age is at
the terminal time T.
12.2.3 Interpretation of the Adjoint Function
An animal at age y at time t, where (t,y) is in D1D2,
will mature at time t-y+h. Its slaughter value at that
time is Q(t-y+h). However, the total feeding and
corralling cost in keeping the animal from its age y until
it matures is given by
Thus, (t,y) represents
the net benefit obtained from having an animal at age
y at time t.
Interpret the optimal control u* in (12.47). Whenever
(t,y) > P(t,y) , we buy more than the goal level
and when (t,y) < P(t,y), we buy less than the goal
level.
12.3 Impulse Control
Example of an oil producer who pumps oil from a
single well.
where 1 is the starting stock of a new oil well.
Figure 12.3: Solution of Equation 12.52
12.3.1 The Oil Driller’s Problem
If t = ti , then
, which means that we have
abandoned the old well and drilled a new well with
stock equal to v(ti).
where P is the unit price of oil and Q is the drilling cost
of drilling a well having an initial stock of 1.
12.3.2 The Maximum Principle for Impulse Optimal
Control
An impulse control variable v , and two associated
functions. The first function is G(x,v,t), which represents
the cost of profit associated with the impulse control.
The sencond function is g(x,v,t), which represents the
instantaneous finite change in the state variable when
the impulse control is applied.
Impluse Hamiltonian function
When t1 = 0 then the equality sign in (vii) should be
replaced by a  sign when i = 1.
Note that condition (vii) involves the partial derivative
of HI with respect to t. Thus, in autonomous problems,
where
condition (vii) means that the
Hamiltonian H is continuous at those times where an
impulse control is applied.
12.3.3 Solution of the Oil Driller’s Problem
Optimal impulse control at t1 is
After the drilling,
(12.70).
which is given by
shown in Figure 12.4, which represent the condition
prior to drilling. Figure 12.4 is
drawn under the assumption that
Figure 12.4: Boundary of No-Drilling Region
From (12.70),
The curve
is drawn in Figure 12.5.
BC of the  curve lies in the no drilling region, which is
above the  curve as indicated in Figure 12.5. The part
AB of the  curve is shown darkened and represents
the drilling curve for the problem. The optimal state
trajectory starts from x(0)=1 and decays exponentially
at rate b until it hits the drilling curve AB at point Y.
Figure 12.5: Drilling Time
Figure 12.6: Value of
This intersection point Y determines the drilling time t1.
12.3.4: Machine Maintenance and Replacement
T = the given terminal or horizon time,
x(t) = the quality of the machine at time t, 0 x  1; a
higher value of x denotes a better quality,
u(t) = the ordinary control variable denoting the rate of
maintenance at time t ; 0 u  G < b/g,
b = the constant rate at which quality deteriorates in
the absence of any maintenance,
g = the maintenance effectiveness coefficient,
 = the production rate per unit time per unit quality of
the machine,
K = the trade-in value per unit quality, i.e., the old
machine provides only a credit against the price of
the new machine and it has no terminal salvage
value,
C = the cost of new machine per unit quality; C > K,
t1 = the replacement time; for simplicity we assume at
most one replacement to be optimal in the given
horizon time; see Section 12.3.3,
 = the replacement variable, 0    1;  represents a
fraction of the old machine replaced by the same
fraction of a new machine. This interpretation will
make sense because we will show that v is either 0
or 1 .
We have assumed that a fraction  of a machine with
quality x has a quality x . Furthermore, we note that
the solution of the state equation will always satisfy
0 x 1, because of the assumption that u  U  b/g.
12.3.5 Application of the Impulse Maximum
Principle
The solution of (12.84) for t1 < t  T is
The switching point
Thus,
is given by solving –1+g=0.
provided the right-hand is in the interval (t1,T];
otherwise set
We can graph the optimal
maintenance control in the interval (t1,T] as in Figure
12.7. Note that this is the optimal maintenance on the
new machine. To find the optimal maintenance on the
old machine, we need to obtain the value of (t) in the
interval (t1,T] .
Figure 12.7: Optimal Maintenance Policy
and compute the time
which makes
In plotting Figure 12.8, we have assumed that 0<(0)
<1. This is certainly the case, if T is not too large so
that
As in the oil driller’s problem, we obtain (t) by using
(12.88). From (12.92), we have v*(t1)=1 and, therefore,
(t1) =K from (12.85) and
from (12.83). Since
gK  1 from (12.96), we have
and, thus, u*(t1)=0 from (12.90). That is, zero
maintenance is optimal on the old machine just before
it is replaced. Since
from (12.97), we have
from Figure 12.7. That is, full maintenance is
optimal on the new machine at the beginning.
Figure 12.8: Replacement Time t1 and Maintenance
Policy
AB represents the replacement curve. The optimal
trajectory x*(t) is shown by CDEFG under the
assumption that t1 > 0 and t1< t2 , where t2 is the
intersection point of curves (t) and (t), as shown in
Figure 12.8.
Figure 12.9 has been drawn for a choice of the
problem parameters such that t1= t2 .
Figure 12.9: The Case t1= t2
Using (t1)=K obtained above and the adjoint equation
(12.84), we have
Using (12.100) in (12.90), we can get u*(t), t[0,t1].
and the switching point
by solving –1+g=0.
If
 0, then the policy of no maintenance is optimal
in the interval [0,t1]. If > 0, the optimal maintenance
policy for the old maintenance is
In plotting Figure 12.7 and 12.8,we have assumed
>0.