3.
Recourse problems
SP with probabilistic constraints as dealt with above is not the only type of stochastic
program. Another type that can be defined deals with what are called recourse problems.
In the simplest model of this type we have two stages:
in the first stage we make a decision
in the second stage we see a realization of the stochastic elements of the problem BUT
are allowed to make further decisions to avoid the constraints of the problem
becoming infeasible.
In other words, in the second stage we have recourse to a further degree of flexibility to
preserve feasibility (but at a cost). Note particularly here that in this second stage the
decisions that we make will be dependent upon the particular realization of the stochastic
elements observed.
To illustrate a simple two-stage recourse model suppose that we have to make a decision
about the amount of a product X to produce. Each unit of X that we make costs us £2. X is
made to meet demand from customers in the next time period. However demand is stochastic,
with a discrete probability distribution: demand = Ds with probability ps (s=1,...,S). Informally
we can think of having S scenarios for possible future demand.
Customer demand must be met. We have the flexibility to buy in the product from an external
supplier to meet observed customer demand but this costs us £3 per unit (i.e. we have
recourse to an additional source of supply if demand exceeds production). How much should
we choose to make now before we know what customer demand is?
Below we show one particular example with S=2 and D1=500, p1=0.6; D2=700, p2=0.4. Note
here that in this example we have looked forward one period into the future in trying to plan
production.
We have to decide how much to produce now before demand is known. If we were to produce
600 (say) then if demand is 500 we are OK, if demand is 700 however we need recourse to an
extra 100 units to meet it.
One way to think of this two-stage model is:
action, make a decision (amount to produce)
observation, observe a realization of the stochastic elements (demand that occurs)
reaction (recourse), further decisions, depending upon the realization observed (extra
production to meet demand if necessary)
You need to be clear about the sequence that applies as you move down the tree in order
to formulate the problem correctly.
We proceed as is usual for modeling a problem, by defining variables. Let: x1 >= 0 be the
number of units of X to produce now (at the first stage)
As we have S scenarios we associate a scenario subscript with the recourse variables at the
second-stage, so let y2s >= 0 be the number of units of X to buy from the external supplier at
the second stage in scenario s when the stochastic realization of the demand is Ds (s=1,...,S).
It is easy to get confused with the subscripts - the order of subscripts is stage, scenario -
where the scenario subscript is dropped if the variable is scenario independent.
Then the constraints to ensure that demand is always satisfied are:
x1 + y2s >= Ds s=1,...,S
Note that we must have >= here, the amount x1 we produce may (depending upon the
particular demand that occurs) exceed customer demand. If we insisted that x1 + y2s = Ds then
as y2s>=0 this would implicitly imply x1<=min[Ds | s=1,...,S].
For the objective function we have a cost 2x1 incurred with certainty and S costs 3y2s, each
incurred with probability ps. Note here that, in practice, only one of these S costs will be
incurred, once a realization of the demand occurs. However before that happens we can only
consider the probability distribution.
It would seem natural to minimize total expected cost, where total expected cost is:
2x1 + SUM[s=1,...,S]ps(3y2s)
Hence our complete simple SP with (linear) recourse is:
minimize 2x1 + SUM[s=1,...,S]ps(3y2s)
subject to
x1 + y2s >= Ds s=1,...,S
x1 >= 0
y2s >= 0 s=1,...,S
Note that this is actually a deterministic program. Note too that we could, if we wish, require
x1 and y2s to be integer.
Let us be clear about what solving this SP will tell us:
we will obtain a value for x1, which is the amount of production we need now
we also obtain a set of values y2s, one for each of the S scenarios of customer demand
Ds (in effect we decide what the optimal decisions are for all possible scenarios)
only one of these S values will be relevant once customer demand is known (i.e. once
we have a realization of stochastic demand), the others will be irrelevant.
To summarize then, in a two-stage SP with recourse we:
make first stage decisions, knowing only the probability distribution for the stochastic
elements
have recourse (at a cost) to second stage variables to ensure constraint feasibility,
these variables being different for different realizations of the stochastic elements
minimize total expected cost, the sum of the known (certain) costs for the first stage
decisions plus the expected cost for the second stage decisions.
Note here that what made our formulation of the problem simple was the fact that the
stochastic element (customer demand) had a discrete distribution. If the stochastic elements
have continuous distributions then the mathematical problems associated with formulation
and solution become formidable
