Model Validation: An Example


We collected data from two human operators, Stan and Oliver, in a driving simulator. For each individual we train a simple HCS model using cascade neural networks. The resulting source and model trajectories are shown below, along with the corresponding similarity measures.


The similarity results confirm two qualitative assessments of the data. First, we observe that the two driving styles are objectively quite different. This fact is reflected in the low similarity measures between one individual's model and the other individual's source and model-generated data. Second, Stan's model is a better reflection of his driving style, than Oliver's model is of his, as reflected in the two respective similarity measures, 0.911 and 0.487. This is indicative that Oliver's sharply discontinuous driving strategy is more difficult to learn by a single cascade network than Stan's calmer approach. Indeed, Oliver's model generates significant oscillatory behavior, of which Oliver himself is not guilty. Finally, we note that the similarity measure serves a dual purpose in the context of human control strategy. It not only serves to validate HCS models, but also can compare and contrast control strategies of different individuals. For this example, it is qualitatively apparent that the driving strategies are quite different. Consequently, the similarity measure evaluates to very small values across individuals.

Back to human control strategy page.


Last updated January 15, 1995 by Michael C. Nechyba