Search

In this document we report the integration and refinement of the different systems developed in work package 2 into a consistent whole that is used in an integrated manner in the PAL system. In the PAL project all developments from the work packages have been integrated in early stages in an iterative manner, providing a running system that can be used by the different stakeholders. As such, all outcomes from WP2 were immediately integrated in the PAL system. This means, for example, that the pal authoring tool (PAL control, Task 2.1) was already integrated in the PAL system when it was ready. In this document we will highlight how the final developments of robot style, action explanation, PAL inform and PAL control were integrated in the final experiment in 2018/19.

In the third design-test cycle of PAL, we refined the prototype based on the outcomes of the previouscycles and additional user feedback sessions. The new protype was improved on its overall usability,goal setting & achievements framework, personalised child-agent interactions, educational content,gamification and the dashboards for parents (PAL Inform). The underpinning knowledge andrationale was recorded in the ontologies and Socio-Cognitive Engineering Tool (SCET; see PALdeliverable D6.13). In an experiment, care with PAL was compared with care as usual for thediabetes self-management of children with T1DM (age 7-14), focusing on (1) determinants ofbehaviour knowledge, awareness and skills), (2) the actual self-management behaviour (i.e.,performing self-management activities and adhering to personalised diabetes goals), and (3) thetreatment outcomes (i.e., glycaemic control). PAL showed to improve children’s knowledge, wheretime of PAL usage correlated positively with this knowledge. In addition, PAL proved to increasethe self-care score (particularly for the younger children), and enhanced the diabetes related qualityof life score of children. Compared to thisPAL prototype version (3.0), children spent their timemore efficiently with an improved version (3.5).

The present report describes the work carried out during the third projectyear regarding PAL's Dissemination activities. It is the summary of threedifferent WP6 Deliverables: Deliverable 6.3 "Website y3", Deliverable 6.7 "Publications and proceedings report y3" and Deliverable 6.11 "Dissemination events promoted y3".

Appendix 1. Examples scenarios. This appendix shows examples of the different scenarios provided to the participants to convey their imagined situation. (DOCX 13 kb)

Scenario-based learning uses interactive scenarios to present the user with situations that need user input to be resolved in order to teach the user the correct behaviour. Traditionally these scenarios would be presented in a rigid order that is linearly increasing in difficulty. Every student however has a different learning rate. Studies have shown that students can lose motivation when challenged too much or too little. This project aims to improve learning efficiency and user satisfaction by adapting the scenario content to the user's skill and knowledge level. A driving instructor application was developed where users are presented with short interactive scenarios in a 3D environment where they take control of the vehicle using a steering wheel and pedals and learn to make correct decisions when driving. The sessions were divided into two different modes, linear or adaptive. In the adaptive mode, the user's performance in the previous scenarios determines which scenario is presented next. Data is collected on user performance, self-efficacy and user-satisfaction.

Recent advances in interpretable Machine Learning (iML) and eXplainable AI (XAI) construct explanations based on the importance of features in classification tasks. However, in a high-dimensional feature space this approach may become unfeasible without restraining the set of important features. We propose to utilize the human tendency to ask questions like "Why this output (the fact) instead of that output (the foil)?" to reduce the number of features to those that play a main role in the asked contrast. Our proposed method utilizes locally trained one-versus-all decision trees to identify the disjoint set of rules that causes the tree to classify data points as the foil and not as the fact. In this study we illustrate this approach on three benchmark classification tasks., Comment: presented at 2018 ICML Workshop on Human Interpretability in Machine Learning (WHI 2018), Stockholm, Sweden