Frank Miller, institutionen för datavetenskap, Linköpings universitet

SEMINARIUM
Datum: onsdag 30 september 2026
Tid: 13:00 - 14:00
Plats: Campus Albano, lärosal 29, hus 4, plan 2

Optimizing model-based incomplete block designs: A comparison of algorithms

Abstract

Because of time limitations or participant burden, the number of treatments in an experimental study may be too large for a single subject. In such cases, incomplete block designs can be used. We consider situations in which a model can be assumed for the relationship under investigation. Based on this model, we define an optimality criterion for selecting the incomplete block design. Identifying the optimal design requires solving a combinatorial optimization problem. Since exact optimization is computationally intractable in this setting, we present local search heuristic algorithms and compare them with existing methods.

We evaluate first- and best-improvement algorithms, simulated annealing, threshold accepting, and two novel algorithms that use directional derivatives to guide exchanges. As discrete analogs of continuous gradient-based methods, these directional derivative algorithms avoid flat regions by prioritizing exchanges associated with large differences in the objective function.

We use item calibration in achievement tests as a comparative example in our broadly applicable framework, evaluating objective values and computational times. The results demonstrate that, for larger problems, the directional derivative algorithms achieve near-optimal solutions significantly faster than simulated annealing and threshold accepting. Their computational efficiency makes these algorithms an appealing approach for practical applications.

This talk is joint work with Jonas Bjermo.

Senast uppdaterad: 2026-09-23

Sidansvarig: Statistiska institutionen