Dealing with uncertainty in test assembly

The recent development of computer technologies enabled test institutes to improve their process of item selection for test administration by means of automated test assembly (ATA). A general framework for ATA consists in adopting mixed-integer programming models which are commonly intended to be solved by commercial solvers. Those softwares, notwithstanding their success in handling most of the basic ATA problems, are not always able to find solutions for highly constrained and large-sized instances. Moreover, all the model coefficients are assumed to be fixed and known, an hypothesis that is not true for the item information functions which are derived from the estimates of item response theory parameters. These restrictions motivated us to find an alternative way to specify and solve ATA models. First, we suggest an application of the chance-constrained (CC) approach (see Charnes and Cooper, 1959) which allows to maximize the α-quantile (usually smaller than 0.05) of the sampling distribution function of the test information function obtained by bootstrapping the calibration process. Secondly, for solving the ATA models, CC or not, we adapted a stochastic meta-heuristic called simulated annealing (SA) proposed by Goffe (1996). This technique can handle large-scale models and non-linear functions. A reformulation of the model by the Lagrangian relaxation helps to find the most feasible/optimal solution and, thanks to the SA, more than one neighbourhood of the space is explored avoiding to be trapped in a local optimum. Several simulations on ATA problems are performed and the solutions are compared to CPLEX 12.8.0 Optimizer, a benchmark solver in the linear programming field. Finally, a real data application shows the potential of our approach in practical situations. The algorithms are coded in the open-source framework Julia. Two packages written in Julia are released for solving the estimation and assembly problems described in this dissertation.