Adaptive Experimental Methods for Evaluating Computational Models of Cognition

描述(申请人提供)：该研究项目的长期目标是提高心理科学中的科学推理能力。这一主题是在认知的计算模型的背景下进行研究的，由于它们的复杂性和相互模仿的程度，在实验上区分这些模型可能非常困难。统计方法(拟合度、Akaike信息准则)已经成为模型评价和选择的主要手段，并在实验中收集数据后才被应用。目前的项目探索了一种新的方法，通过开发相应的统计方法来改进推理，这些方法在设计实验时应用于实验的前端。在这种被称为自适应设计优化(ADO)的方法中，实验被分成一系列的微型实验。每个迷你实验的设计都会根据前一个迷你实验中的表现进行更新。设计值的选择是由复杂的搜索算法决定的，该算法不断向感兴趣的模型施压，以适应越来越具挑战性的数据点，直到一个模型脱颖而出。该方法的自适应性质确保了设计在整个测试过程中都是最佳的，从而使实验结果的信息量最大化。此外，对优化设计的关注同时确保了实验的高效率(例如，较少的试验和参与者)。该提案的三个具体目标是：(1)开发ADO，使其适用于学科中的广泛问题(例如，各种实验设计，不同的建模目标)；(2)改进ADO算法，使其能够用于实时实验；(3)开发基于网络的资源，使研究人员能够了解和利用该方法。这三个目标的实现意在为研究人员提供一种可以加速科学发现的新技术。

DESCRIPTION (provided by applicant): The long-term goal of this program of research is to improve scientific inference in psychological science. The topic is investigated in the context of computational models of cognition, which can be extremely difficult to distinguish experimentally because of their complexity and the extent to which they mimic each other. Statistical methods (goodness-of-fit, Akaike Information Criterion) have been the dominant means of model evaluation and selection, and are applied after data have been collected in an experiment. The current project explores a new approach to improving inference by developing corresponding statistical methods that are applied on the front-end of an experiment, while the experiment is being designed. In this approach, dubbed adaptive design optimization (ADO), an experiment is divided into a series of mini-experiments. The design of each mini-experiment is updated based on performance in the preceding mini-experiment. The choice of design values is dictated by a sophisticated search algorithm that constantly pressures the models of interest to fit more and more challenging data points until one model emerges as superior. The adaptive nature of the methodology ensures the design is optimal throughout the testing session, and thereby maximizes the informativeness of the experimental results. Furthermore, the focus on optimizing the design simultaneously ensures that the experiment is highly efficient (e.g., fewer trials and participants). The three specific aims of the proposal are to (1) develop ADO so that it is applicable to a broad range of problems (e.g., various experimental designs, different modeling goals) in the discipline; (2) improve the ADO algorithm so that it can be used in real-time experiments; (3) develop web-based resources to enable researchers to learn about and take advantage of the methodology. The achievement of these three goals is intended to provide researchers with a new technology that can accelerate scientific discovery.