Figure 2.4. A simple instance for a bilevel hazmat transportation problem
2.2
BLP relaxations
2.2.1
Relaxation via removal of constraints
The different impact of upper level and lower level constraints on the bilevel formulation has to be
taken into account when we want to relax a BLP by dropping a subset of constraints. Unlike classical
mathematical programming models, the new formulation obtained removing one or more constraints
does not necessarily provide a valid relaxation. Indeed, this depends on the choice of the constraints
we remove. Let us consider the following problem:
min
x,y−x + 3y
s.t.
y ≥ 1
y ∈ argmin
yy
s.t.
x + y ≤ 4
x + y ≥ 2
x − y ≥ −1
x − y ≤ 1
The optimal solution of the BLP is (2, 1), which is vertex C of S, and the optimal value is 1, see Figure
2.5(a). If the upper level constraints y ≥ 1 is dropped, as depicted in Figure 2.5(b), the inducible
region is wider, and the new optimal solution is vertex E, (1.5, 0.5), and the objective function’s value
is 0. Finally, if the follower’s constraint x − y ≤ 1 is removed, the follower’s feasible set is larger, but
the new rational solutions of the follower violate the upper level constraint. The inducible region is
reduced to segment A–B and the optimal solution is vertex B, (1, 1), which gains an objective func-
tion equals 2. Hence, in the second reformulation represented in Figure 2.5(c), dropping a follower’s
constraint we do not obtain a valid relaxation and the optimal solution is the worst out of the three cases.
(a) Inducible region of the original problem (b) Removal of an upper level constraint
(c) Removal of a lower level constraint
2.2 BLP relaxations 27
This result, which is apparently unclear, can be easily understood if we think to the inherent meaning
of bilevel programming problems. The general framework of bilevel programming is the presence of
two non cooperating decision makers. Dropping a lower level constraint means that the set of possible
choices for the follower is increased and thus he acquires more contractual power towards the leader.
This implies that the solutions space of the leader may reduce and the value of the optimal solution
may get worse. On the contrary, if an upper level constraint is removed, the solutions space of the
follower does not change, but a larger set of rational solutions may be considered acceptable by the
leader. Thus the value of the optimal solution can not be worse and this represent a valid relaxation.
The following proposition sums up the above mentioned results.
Proposition 2. A BLP can be correctly relaxed dropping a subset of constraints, if and only if these
are constraints under control of the leader.
2.2.2
Single level relaxation
One of the most used and immediate relaxation of a BLP is the so called single level relaxation. It
consists of dropping the follower’s objective function, turning the bilevel structure into a single level
one and assuming that only one decision maker is involved; from a mathematical point of view the
leader’s feasible set is fully defined by a set of constraints and it is not necessary to solve an inner
problem to check feasibility of a solution. The feasible set S does not change and the problem consists
of solving the leader’s objective function on S.
It is immediate to understand that the single level version of a BLP is a valid relaxation, as a
solution that is bilevel–feasible for the BLP is also feasible for the single level problem, but the contrary
may not be true.
There is a special case in which the optimal solution of a BLP and of the optimal solution of its
single level version coincides. For the sake of simplicity, let us assume that there are no upper level
constraints. If the following holds
∇
yF (x, y) = ∇f (y)
it means that the leader’s and the follower’s objective function have the same verse, thus minimizing
(or maximizing) F (x, y), f (y) is minimized (or maximized) as well. In this case, the optimal solution
found solving the single level formulation is also optimal for the BLP.
This result may not hold if we introduce an upper level constraint which depends on the upper
level variables x. We now prove the following two theorems, which are necessary and sufficient
conditions under which the optimal solution of the single level relaxation of a BLP is also optimal for
the original problem. According to these conditions, it is easy to know whether the optimal solution of
the relaxation is bilevel–feasible, without solving the inner problem. We assume that S is compact and
non empty.
Theorem 10. If the optimal solution (¯x, ¯y) of the single level relaxation of a BLP is optimal for the
original BLP, then∇
yF (x, y) = ∇f (y) and among the active constraints at (¯x, ¯y) there is at least
one constraint under control of the follower.
Proof. Let us assume that (¯x, ¯y) is a vertex of S in which the only active constraints are under control
of the leader. Let us define S
0= S \ {Cx + Dy ≤ e}. From the previous sections we know that
S ⊆ S
0and that solving the single level relaxation of BLP on S
0two cases may occur: a) we found a
solution (x
0, y
0) on the boundary of S
0, b) the single level relaxation does not admit a finite solution,
i.e. S
0is unbounded. In case a), if (x
0, y
0) ≡ (¯x, ¯y), it means that there is at least an active constraint at
(¯x, ¯y) which is not under control of the leader and this contradicts our initial assumption. It follows that
(¯x, ¯y) is an internal point of S
0, thus there exists a rational solution (¯x, y
∗) such that f (y
∗) < f (¯y). In
case b), once again, (¯x, ¯y) is an internal point of S
0and ¯y is not a rational solution at ¯x. In both cases
(¯x, ¯y) is not bilevel–feasible, hence the proof.
2
Theorem 11. Given the optimal solution (¯x, ¯y) of the single level relaxation of a BLP, if ∇
yF (x, y) =
∇f (y) and among the active constraints at (¯x, ¯y) there are no upper level constraints, (¯x, ¯y) is also
the optimal solution of BLP.
Proof. If there are no upper level constraints active at (¯x, ¯y), it means that if we solve the single level
relaxation on S
0, defined as above, the optimal solution found is always (¯x, ¯y). Thus (¯x, ¯y) is a rational
solution and satisfies the upper level constraints, hence it is bilevel–feasible and optimal for the BLP,
which completes the proof.
2
In Figures 2.6(a)-(d) there is a graphic explanation of Theorems 10 and 11. In each figure we reported
both the feasible set S and some additional upper level constraints.
In Figure 2.6(a) an upper level constraint intersects the inducible region and solution A is a vertex of
S in which a lower level constraint is active; A is bilevel–feasible and is the optimal solution of the
BLP. In Figure 2.6(b), although a lower level constraint is active at A, unlike the previous case the
solution is not bilevel–feasible and the optimal solution of the BLP is vertex B: this shows that the
condition stated in Theorem 10 is only necessary and not sufficient. In Figure 2.6(c) an upper level
constraint intersects the inducible region but this does not change the optimal solution of the single
level relaxation problem and Theorem 11 holds. Finally, in Figure 2.6(d) all the constraints active at
vertex A are upper level constraints and by Theorem 11 the solution is not optimal for the BLP.
In document
Integer Bilevel Linear Programming Problems: New Results and Applications
(Page 43-46)