[Python] Ortools can not evaluate LinearExpr edge case

I encounter a specific edge case in a CP model. All my variables are int based but the objective which is float based.

In a lot of usage of my model, there is no issue doing this. But in one specific instanciantion, I encounter an error while evaluating the objective.

An error is raised in the evaluation function :

# The librairie function
def evaluate_linear_expr(
expression: LinearExprT, solution: cp_model_pb2.CpSolverResponse
) -> int:
"""Evaluate a linear expression against a solution."""
if isinstance(expression, IntegralTypes):
return int(expression)
if not isinstance(expression, LinearExpr):
raise TypeError("Cannot interpret %s as a linear expression." % expression)

value = 0
to_process = [(expression, 1)]
while to_process:
expr, coeff = to_process.pop()
if isinstance(expr, IntegralTypes):
value += int(expr) * coeff
elif isinstance(expr, _ProductCst):
to_process.append((expr.expression(), coeff * expr.coefficient()))
elif isinstance(expr, _Sum):
to_process.append((expr.left(), coeff))
to_process.append((expr.right(), coeff))
elif isinstance(expr, _SumArray):
for e in expr.expressions():
to_process.append((e, coeff))
value += expr.constant() * coeff
elif isinstance(expr, _WeightedSum):
for e, c in zip(expr.expressions(), expr.coefficients()):
to_process.append((e, coeff * c))
value += expr.constant() * coeff
elif isinstance(expr, IntVar):
value += coeff * solution.solution[expr.index]
elif isinstance(expr, _NotBooleanVariable):
value += coeff * (1 - solution.solution[expr.negated().index])
else:
# It raises here !
raise TypeError(f"Cannot interpret {expr} as a linear expression.")

return value


The error: TypeError: Cannot interpret 4000000.0 as a linear expression.

Here is what I can see in debug:

> expression
(7.209805335255949e-08 * (((((((3000 * (VAR300 + (quotient + has_remainder))) + (93000 * VAR879)) + (60000 * VAR300)) + ((1500 * VAR564) + (25000 * VAR780))) + ((2000 * VAR578) + (140000 * VAR781))) + ((9570.300000000001 * VAR832) + (40000 * VAR882))) + 4000000.0))

> type(expression)
<class 'ortools.sat.python.cp_model._ProductCst'>

> type(expr)
<class 'float'>



I feel like it should be handled by the librairie... or I am doing something wrong ?

Versions Used: Python 3.9.0 ; ortools 9.11.4210

Update I figured out that the only difference between the working instances and this one is the inclusion of a constant float offset, which the evaluation function is unable to process. I found it inconsistant since, in any case, my objective is a floating value and is correctly handled without this constant.

Continue reading...