2.4.3. Finding the Most Promising Coaching Determinants

Once the coaching domain is selected, the system investigates on which personal determinants the coaching messages should focus to yield the most promising effect on the desired behavior. These behavioral determinants are personal psychological concepts that govern the engagement in healthy behavior. The system contains a large collection of coaching messages, with subsets that each target one of the personal determinants. The messages are based on established behavior change techniques, such as prompting barrier identification, providing information on consequences, and prompting goal setting [7].

In order to determine what messages are most likely to positively affect the user's behavior, the effects of improving each one of the personal determinants are estimated based on simulations of a dynamic computational model. This model is a formalization of the dynamics between these personal determinants and the behavior, where each of the concepts is represented by a numerical value in range [0, 1]. The model is mainly based on the social cognitive theory, that describes the reasons why people fail or succeed to exhibit some desired (health) behavior from both social and cognitive determinants [14,15]. Other theories (e.g., self-regulation theory and health action process approach) and literature were consulted to extend the model to incorporate more relevant aspects. The model that was implemented in the Active2Gether system is an adaptation of the computational model presented in [16]. Revisions between the two versions of the model were motivated by a decrease in conceptual detail and computational complexity of the model, and by suggestions of experts in the domain of behavior change. The resulting model contains determinants such as intentions, self-efficacy and outcome expectations. The values of all parameters in the model can be adjusted by the modeler. In the current implementation, the parameters were chosen based on correlations between the concepts found in literature [17–19]. We performed simulations to find values that keep the ratio between the parameters in accordance with empirical findings [20]. A more detailed description and a preliminary validation of this model on empirical data, which showed promising results, can be found in [21]. Figure 5 shows a graphical representation of the computational model.

**Figure 5.** Graphical representation of the computational model.

The simulation process starts with estimating the current states of the personal determinants by means of short questions via the app. The resulting values are used as input for the computational model. To simulate the effect of targeting one of the determinants, one of the values obtained from the app questionnaire is increased according to the hypothesized effect of sending coaching messages about this determinant. Then, the computational model simulates the dynamics between the determinants and estimates the effect on the behavior. By running simulations for each possible targeted determinant, a list of determinants is constructed, ordered by the most promising effect on the behavior variable. This order is taken into account when selecting coaching messages to the user. As with the selection of a coaching domain and goal, the simulation cycle is repeated weekly, in order to tailor to the user's strongest psychological needs at all times.

In contrast to the relative evaluation of the user's behavior for suggesting a coaching domain, this part of the reasoning engine does not tailor the intervention based on information about the user's environment, but rather on information about his/her motivational state of mind. This way, the users will receive support on the aspects that are relevant to their motivation and behavior.
