Analyzing Complex Stochastic Models using Approximate Bayesian Computation

描述(由申请人提供)：采用贝叶斯框架而不是经典技术有许多理论和实践原因。最近，许多心理学模型受益于贝叶斯分析，特别是它们受益于提供多个层面信息的层次分析。然而，能够享受贝叶斯分析好处的模型类别一直局限于具有易于处理的似然函数的模型。这些模型通常被称为基于模拟的模型，由于其复杂或难以处理的似然函数，贝叶斯分析是不可能的。然而，一种名为近似贝叶斯计算(ABC)的新技术允许研究人员通过模拟模型来规避可能性评估。我们提出的研究将把ABC的应用扩展到计算神经科学的复杂、随机模型。对于这些模型，我们将有兴趣拟合模型的分层和有限混合版本，以检查个体差异，并探索实验设计优化在模型选择中的作用。 公共卫生相关性：计算神经科学的随机复杂模型的分析将促进我们对指导感兴趣行为的认知系统中潜在机制的理解。引入贝叶斯无似然推理的一般框架将对任何使用计算模型的领域产生广泛的影响。这将在计算神经科学中引起特别的兴趣，在计算神经科学中，人类行为的模型通常非常复杂，具有复杂的似然函数。

DESCRIPTION (provided by applicant): There are a number of theoretical and practical reasons to adopt the Bayesian framework over classical techniques. Recently, many psychological models have benefited from Bayesian analyses, and in particular, they have benefited from hierarchical analyses that provide information on multiple levels. However, the class of models that are capable of enjoying the benefits of Bayesian analyses has been limited to models that possess tractable likelihood functions. These models are typically referred to as "simulation-based" models, and as a result of their complicated or intractable likelihood functions, Bayesian analyses are not possible. However, a new technique, called approximate Bayesian computation (ABC), allows researchers to circumvent the evaluation of the likelihood by simulating the model. Our proposed research will extend the application of ABC to complex, stochastic models of computational neuroscience. For these models, we will be interested in fitting hierarchical and finite mixture versions of the models to examine individual differences and explore the role of experimental design optimization in model selection. PUBLIC HEALTH RELEVANCE: The analysis of stochastic complex models of computational neuroscience will advance our understanding of the underlying mechanisms in the cognitive system guiding the behavior of interest. Introducing the general framework for Bayesian likelihood-free inference will have a broad impact on any field employing computational models. This will be of particular interest in computational neuroscience where models of human behavior are often very sophisticated and have complex likelihood functions.