This paper proposes an optimal experiment design approach for parameter estimation in a bounded-error context. In the design phase, this approach does not require specifying any reference value of the vector of parameters, contrary to the state-of-the-art techniques. Two variants of the problem are considered, depending on whether one is interested in the experiment minimizing the volume (DB-SM optimal experiment) or the sum of the edges of the smallest box containing the set estimate (A-SM optimal experiment). To obtain the DB-SM or A-SM optimal experiments, an outer approximation of the smallest box containing the set estimates has to be evaluated, independently of the value of the true value of the parameter vector and of the measurement noise. For that purpose, existentially quantified linear programs have to be solved. One shows that these quantified linear programs can be transformed into several classical linear programs. Illustrations on a simple exponential example show that the DB-SM or A-SM optimal experiments lead to smaller set estimates in the worst case, compared to experiments designed with state-of-the-art techniques.