This paper investigates improvements to modeling other agents based on observed situation-action pairs and the nearest-neighbor rule, which suffers when dealing with very few or very many observations. Stereotype models allow for good predictions of a modeled agent s behavior even after few observations. To handle any adverse effects of stereotyping, periodic reevaluation of the chosen stereotype and the potential to switch between different stereotypes aids in dealing with very similar, but not identical, stereotypes. Also, periodic reevaluation and the potential to switch from a chosen stereotype to the original observation based modeling method aids in dealing with agents that do not conform to any stereotype. Finally, compactification of observation keeps the application of the original modeling method efficient by reducing comparisons within the nearest neighbor rule.

Our experiments within the OLEMAS system show that stereotyping significantly improves cases where just using the original modeling performs poorly. In addition, reevaluation and switching fortify stereotyping against the potential risk of using an incorrect stereotype. In many cases, compactification both increases the efficiency of using observed situation-action pairs and improves predictions by filtering out misleading observation. At times however, this filtering comes at the cost of prediction accuracy by removing relevant observations.