Perception and Action: Ideal Observers and Actors

描述（由申请人提供）：人类在决策或行动的所有方面都可能是随机的情况下做出决策和执行行动。基于感官信息的行动计划有五个组成部分。首先，主体具有关于环境状态的先验信息，包括附近物体和主体自身的当前位置和速度；这个信息几乎肯定是不完整的，可以概括为可能世界状态的概率分布。其次，受试者对环境的当前状态有感官输入，由于物理和神经噪声源，这种输入也将是不确定的。第三，受试者将这两种信息来源结合起来，并决定一个预期的动作（按下按钮、伸手运动、眼球运动或更复杂的运动计划，包括对潜在的后续感官输入的反应）。第四，由于电机噪声，产生的动作可能与预期的动作不同。最后，所产生的动作与当前环境的相互作用会给主体带来一个结果（损失或收益），这个结果也可能是随机的。由于所有这些随机成分，视觉任务和运动计划都需要计算，这相当于在风险下决策所需的计算。在我们最近的工作中，我们描述了人类在视觉运动任务中接近最佳的情况，因为他们最大化了预期收益，以及人类行为次优的其他情况。我们建议通过实验来更好地理解人类在视觉和视觉运动任务中的行为本质。我们继续使用具有实验者指定的奖惩结构的任务，以便我们可以将行为与最大化预期收益的最佳策略进行比较。我们提出了以下问题，并提出了实验来解决每个问题：(1)任务不确定性的哪些方面被人类观察者很好地估计并最佳地用于选择运动计划？我们将确定人类是否在风险下最佳规划运动，因为感觉输入和/或运动输出通过各种方式变得更嘈杂。(2)运动计划对视觉运动任务不同组成部分变化的反应有多灵活？我们将测量视觉运动任务的学习进展，其中先验概率、运动结果或回报是不确定的，并且随着时间的推移而变化。(3)在日常生活中，需要对视觉目标的发现、辨别和搜索来引导人们朝着这些目标行动，作为获得后期奖励的一种手段。在这里，我们要问的是，当涉及明确定义的得失时，人类在典型的视觉任务中是否处于最佳状态。我们将通过将人类表现与理想观察者模型进行比较，确定人类在视觉检测、识别和搜索任务中的表现是否最佳，理想观察者模型在不对称收益情况下最大化预期收益。拟议的工作通过描述与做出感知决策或使用感觉信息控制运动有关的神经机制，对公共卫生有益。我们展示了如何优化决策和运动计划必须考虑到先验知识，视觉信息的不确定性，运动反应的可变性和未知的或不断变化的回报。各种医疗条件可以影响视觉信息的可靠性（例如，白内障，弱视等）和运动输出的质量（例如，帕金森病，中风）。拟议的研究将提高我们对最佳感知决策或运动计划的理解，因此可以帮助设计康复计划，当感觉输入或运动输出被疾病或其他与健康相关的条件破坏时（偏差，增益和/或可变性的变化）。

DESCRIPTION (provided by applicant): Humans make decisions and perform actions in situations in which all aspects of the decision or action are potentially stochastic. There are five components to the planning of an action based on sensory information. First, the subject has prior information about the state of the environment including the current positions and velocities of nearby objects and of the subject's own body; this information is almost certainly incomplete, and can be summarized as a probability distribution across possible world states. Second, the subject has sensory input about the current state of the environment, and that input will also be uncertain due to physical and neural noise sources. Third, the subject combines these two sources of information and decides on an intended action (button press, reaching movement, eye movement, or a more complex movement plan that includes responses to potential subsequent sensory inputs). Fourth, the resulting action can differ from the intended one due to motor noise. Finally, the interaction of the resulting action with the current environment leads to a consequence (a loss or gain) for the subject, and this consequence may be random as well. As a result of all these stochastic components, both visual tasks and movement planning require a calculation that is equivalent to that required for decision-making under risk. In our recent work, we have delineated situations in which humans are nearly optimal in visuo-motor tasks in that they maximize expected gain, and other circumstances in which human behavior is suboptimal. We propose experiments to better understand the nature of human behavior in visual and visuo-motor tasks. We continue to use tasks with an experimenter-specified reward/penalty structure so that we may compare behavior with the optimal strategy that maximizes expected gain. We ask the following questions and propose experiments to address each: (1) What aspects of task uncertainty are estimated well by human observers and used optimally to select a movement plan? We will determine whether humans optimally plan movements under risk as sensory input and/or motor output is made noisier by a variety of means. (2) How flexible is movement planning in response to changes in different components of a visuo-motor task? We will measure the progress of learning in visuo-motor tasks in which prior probabilities, motor outcome or payoff are uncertain and changing over time. (3) In daily life, detection, discrimination and search for visual targets are required to guide action toward those targets as a means of obtaining later rewards. Here, we ask if humans are optimal in typical visual tasks when clearly defined gains and losses are involved. We will determine whether human performance in visual detection, discrimination and search tasks is optimal by comparing human performance to ideal-observer models that maximize expected gain in situations with asymmetric payoffs. PUBLIC HEALTH RELEVANCE The proposed work benefits public health by characterizing the neural mechanisms that are involved with making perceptual decisions or using sensory information to control movements. We show how optimal decisions and movement plans must take into account prior knowledge, the uncertainty of visual information, the variability of motor response and unknown or changing payoffs. A variety of medical conditions can impact both the reliability of visual information (e.g., cataract, amblyopia, etc.) and the quality of motor output (e.g., Parkinson's disease, stroke). The proposed research will improve our understanding of what is meant by an optimal perceptual decision or movement plan, and thus can serve to help in the design of rehabilitative plans when sensory input or motor output is disrupted (change in bias, gain and/or variability) by disease or other health-related conditions.