Data and analysis code for the Masters thesis titled "Dyadic Physical Activity Planning with a Virtual Coach: Using Reinforcement Learning to Select Persuasive Strategies"

This is the data and analysis code for the Masters thesis titled "Dyadic Physical Activity Planning with a Virtual Coach: Using Reinforcement Learning to Select Persuasive Strategies." In this thesis, we used data gathered through an observational study with a virtual coach to build a reinforcement learning model for persuading people to commit to plans for walking. Once the model was created, it was analysed using the code provided in this dataset, including simulations of how the model would perform in a real-life setting.<br><strong>Study</strong>The study was conducted on the online crowdsourcing platform Prolific in June and July 2023. The Human Research Ethics Committee of Delft University of Technology granted ethical approval for the research (Letter of Approval number: 3089). 114 participants participated in a conversation with the virtual coach named Jamie. During these conversations, the virtual coach measured people's confidence in following the plan, their perceived usefulness of planning, and their attitude towards planning. Then, the virtual coach used a persuasive strategy to try to change these aspects of a person's state, after which it measured the three aspects again. This repeated until at least two persuasive strategies were used, and until the three aspects were high enough to satisfy pre-set heuristics for when the person is likely to commit to the plan. When these conditions were met, the virtual coach measured the reward signal, composed of satisfaction with the conversation, commitment to the fist two weeks and to the whole plan, and confidence in reaching the goal.<br><strong>Data</strong>We provide data on:<br>participant characteristics (e.g., age, gender, stage of change for becoming physically active),the state-action-next state-reward samples used to create the reinforcement learning model,free-text responses to the persuasive strategies used by the virtual coach,free-text responses about what people found motivating and demotivating in motivational messages.<br>The file OSF pre-registration explains in detail how each variable was measured.<br><strong>Analysis</strong>We provide code to reproduce our analysis with Docker. For more information on this, please refer to the README-file in this repository.<br>In the case of questions, please contact Nele Albers ([e-mail]) or Willem-Paul Brinkman ([e-mail]).

本仓库包含题为《虚拟教练辅助双人运动规划：基于强化学习的说服策略选择》的硕士学位论文相关数据与分析代码。本研究依托与虚拟教练开展的观察性研究所收集的数据，构建了用于说服用户承诺步行计划的强化学习（Reinforcement Learning）模型。模型构建完成后，本数据集附带的代码用于对该模型开展分析，包括模拟模型在真实场景中的表现。<br><strong>研究概况</strong><br>本研究于2023年6月至7月在众包在线平台Prolific上开展。代尔夫特理工大学人类研究伦理委员会为本次研究授予了伦理批准（批准函编号：3089）。共有114名参与者与名为Jamie的虚拟教练进行对话。对话过程中，虚拟教练会采集用户对执行计划的信心、对规划行为的感知有用性，以及对规划的态度三项指标。随后，虚拟教练将运用某一说服策略尝试调整用户上述三项状态，并在调整后再次测量这三项指标。该流程将重复执行，直至至少使用了两种说服策略，且三项指标达到预设启发式阈值（即用户大概率会承诺执行计划的临界值）。当上述条件达成时，虚拟教练将采集奖励信号，该信号由对话满意度、对前两周计划的承诺度、对整体计划的承诺度，以及达成目标的信心共同构成。<br><strong>数据集内容</strong><br>本数据集提供以下数据：<br>• 参与者特征（如年龄、性别、参与体育活动的行为改变阶段）<br>• 用于构建强化学习模型的状态-动作-下一状态-奖励样本集<br>• 用户针对虚拟教练所采用说服策略的开放式文本反馈<br>• 用户关于激励信息中哪些内容具有激励性或反激励性的开放式文本反馈。<br>OSF预注册文件详细说明了各变量的采集方式。<br><strong>分析代码</strong><br>本数据集附带可通过Docker复现研究分析的代码。如需了解更多相关信息，请参阅本仓库中的README文件。<br>如有任何疑问，请联系Nele Albers（邮箱：[e-mail]）或Willem-Paul Brinkman（邮箱：[e-mail]）。