Circuit mechanisms of arbitration between distinct reinforcement learning systems

项目总结 动物可以在新环境中表现出目标导向的行为，尽管经验有限 和他们在一起。大脑如何做出和使用关于潜在统计数据和 环境的生成性结构来指导行为？强化学习领域指的是 将这一能力归结为“基于模型”的推理，这意味着它依赖于 世界的结构。关键是，这个内部模型可以用来灵活地估计最好的 在没有直接经验的情况下，通过心理模拟或计划采取行动。相比之下，在“无模型”中 强化学习，智能体根据直接经验选择最好的动作，不需要 明确了解任务或环境的基本顺序转换结构。 基于模型的机制和无模型的机制在大脑中共存，并由不同的 电路，尽管大脑在这些之间进行仲裁的神经电路机制 决策系统仍不得而知。理论和行为研究表明，人类 大脑使用的系统产生的价值估计的不确定性最低。侧向 眼眶额叶皮质(LOFC)是进行仲裁的有力候选者，因为虽然它是 与基于模型的推理有关，例如，通过启用有关隐藏任务的推理 在美国，它位于背侧纹状体的上游，这对基于模型和模型- 自由决策。有趣的是，我们发现LOFC神经元只投射到 背外侧纹状体(DLS)，一个对无模型行为至关重要的区域，而不是背内侧 纹状体(DMS)，这是基于模型的行为的关键。我们假设这一预测 LOFC中的特定神经电路通过抑制无模型在这些系统之间进行仲裁 系统。 我将使用最先进的病毒、电生理和计算方法来 确定投射DLS的LOFC神经元是否调解基于不确定性的仲裁 决策系统(目标1)，并描述支持以下内容的基本电路逻辑 仲裁(目标2)。通过光遗传学标记DLS投射的LOFC神经元，I将选择性地 在监控大鼠在任务中使用的行为策略的同时，表征和干扰它们的活动 具有潜在的结构。为了确定仲裁是如何在背侧纹状体中实例化的，我将 在记录不同遗传细胞类型时，光遗传学激活OFC→延髓神经元 体内和体外的纹状体。我们预测OFC→DLS神经元支持基于模型的 通过激活抑制性中间神经元来抑制DLS和无模型系统的行为。

PROJECT SUMMARY Animals can exhibit goal-directed behaviors in novel environments, despite limited experience with them. How does the brain make and use inferences about the underlying statistics and generative structure of environments to guide behavior? The field of reinforcement learning refers to this capacity as “model-based” reasoning, meaning that it relies on an internal model of the structure of the world. Critically, this internal model can be used to flexibly estimate the best actions by mental simulation or planning, without direct experience. In contrast, in “model-free” reinforcement learning, an agent chooses the best action based on direct experience, without explicit knowledge of the underlying sequential transition structure of a task or environment. Model-based and model-free mechanisms coexist in the brain and are mediated by distinct circuits, although the neural circuit mechanisms by which the brain arbitrates between these decision systems remains unknown. Theoretical and behavioral studies suggest that human brains use the system that yields value estimates with the lowest uncertainty. The lateral orbitofrontal cortex (lOFC) is a compelling candidate to perform arbitration because while it is implicated in model-based reasoning, for instance by enabling inferences about hidden task states, it lies upstream of the dorsal striatum, which is critical for both model-based and model- free decision making. Intriguingly, we have found that lOFC neurons project exclusively to the dorsolateral striatum (DLS), a region critical for model-free behavior, and not the dorsomedial striatum (DMS), which is critical for model-based behavior. We hypothesize that projection specific neural circuits in lOFC arbitrate between these systems by suppressing the model-free system. I will use state-of-the-art viral, electrophysiological, and computational methods to determine whether DLS-projecting lOFC neurons mediate uncertainty-based arbitration between decision-making systems (Aim 1) and characterize the underlying circuit logic that supports arbitration (Aim 2). By optogenetically tagging DLS-projecting lOFC neurons I will selectively characterize and perturb their activity while monitoring the behavioral strategy rats use in a task with latent structure. To determine how arbitration is instantiated in the dorsal striatum I will optogenetically activate OFC→DLS neurons while recording from different genetic cell types in the striatum, in vivo and in vitro. We predict that OFC→DLS neurons enable model-based behavior by activating inhibitory interneurons to suppress the DLS and the model-free system.