THESIS
1999
Abstract
Based on the adaptive stochastic control approach, we herein propose an alternative item selection procedure for computerized adaptive testing (CAT) to measure a subject's ability parameter θ with greater efficiency. Starting with the identification to a special class of multi-armed bandit problem, our goal is to maximize the total Fisher information under irreversible adaptive allocation of item difficulty levels. We first prove the short fall from complete information total expected reward has a lower bound. Next, we introduce a class of irreversible adaptive allocation rules based on Kullback-Leibler information. The rules are shown to be asymptotically efficient in the sense of achieving the lower bound as the number of administrated items N → ∞. Simulation results show that they al...[
Read more ]
Based on the adaptive stochastic control approach, we herein propose an alternative item selection procedure for computerized adaptive testing (CAT) to measure a subject's ability parameter θ with greater efficiency. Starting with the identification to a special class of multi-armed bandit problem, our goal is to maximize the total Fisher information under irreversible adaptive allocation of item difficulty levels. We first prove the short fall from complete information total expected reward has a lower bound. Next, we introduce a class of irreversible adaptive allocation rules based on Kullback-Leibler information. The rules are shown to be asymptotically efficient in the sense of achieving the lower bound as the number of administrated items N → ∞. Simulation results show that they also perform very well for moderate values of N.
Post a Comment