*3.4. Model-Based Reasoning*

Some of the design choices in the personalization process described in Section 2.4 seem successful, but we identified opportunities for improvement in others. Our findings on these elements of the Active2Gether system are described below.

Based on anecdotic feedback, we conclude that determining the user's awareness phase by comparing their actual behavior to their perception works well. Acknowledging the users' awareness of the need to change is a useful way of tailoring the coaching messages.

In contrast, the suggestion of a coaching domain could be improved. The current approach is not very flexible, as the scores for active transport and stair walking are based on the characteristics of the significant locations that were identified via the intake questionnaire. Any physical activity related to these domains on other locations is ignored during the evaluation of the user's behavior. Since that activity is not taken into account for either the actual behavior or the "ideal" behavior, this simplification should not distort the behavior scores. However, it is recommended to also take behavior on (or during transit to) other locations into account, to get a more complete picture of the user's behavior. This could be achieved by using more adaptive behavior evaluation algorithms, which learn the user's potential or ideal from past behavior, possibly in combination with other (web) sources.

Part of the selection of coaching messages is based on the simulations of a computational model. Although a preliminary validation of the model showed very promising results [21], the added value of the model in predicting the most effective coaching determinants still has to be evaluated. In theory, an adaptive approach can be used to learn the effect of specific (sets of) messages on a person's behavior, which might lead to better suggestions for coaching determinants. The outcome of evaluating the model could for example lead to the decision to use personal and adaptive parameters in the computational model, or to take an entirely different approach (e.g., machine learning techniques).
