CRI: Navigation and the Hippocampus: Computational Models

像老鼠这样能够以复杂的方式操纵空间的动物，有能力根据来自不同模式的大量线索，获取有关其环境布局和内容的大量信息。他们还根据自己不断变化的目标，规划行动，以优化这些环境的使用。这些任务远远超出了目前的机器人技术。人们开始了解老鼠这些能力背后的神经基础。老鼠的海马体是一个解剖学和生理学研究已久的结构，已知其包含的细胞在熟悉的环境中可靠地报告其位置。同时来自海马体多个细胞的记录开始显示这些地图是如何在经历中形成的。我们将结合脑科学中使用的三种关键方法——动物行为、海马体中大脑活动的神经记录和计算机建模，以了解大鼠用于建立环境模型和构建最佳计划的自适应优化控制策略。导航是自主智能体的一个关键问题。这项工作的结果将是大脑功能的模型，利用在空间任务中涉及学习、记忆和控制的生物结构的能力。这些模型将能够整合多感官信息并实现优化控制。基于动物行为理论的类似方法已经广泛应用于各种复杂的组合问题，如调度。我们将采用的用于理解自然导航的强化学习方法应该在智能机器人系统的开发中具有直接的工程应用。

IBN: 9634339 PI: Dayan Animals such as rats that manipulate space in sophisticated ways have the capacity to acquire substantial information about the layout and contents of their environments, based on a large variety of cues from different modalities. They also plan their actions to optimise the use of these environments, based on their changing goals. These tasks are far beyond current robot technology. The neural basis underlying these abilities of rats is starting to be understood. The hippocampus of a rat, a structure whose anatomy and physiology have long been studied, is known to contain cells that report reliably on its location in a familiar environment. Simultaneous recordings from multiple cells in the hippocampus are starting to show how these maps are formed during experience. We will combine three key approaches used in brain science - animal behavior, neural recording of brain activity in the hippocampus, and computer modeling to understand the adaptive optimising control strategies that rats use to build models of their environments and to construct optimal plans. Navigation is a key problem for autonomous agents. The result of this work will be models of brain function which exploit the capacities of the biological structures involved in learning, memory and control in spatial tasks. These models will be capable of integrating poly-sensory information and implementing optimising control. Similar approaches based on theories of animal behavior are already finding wide application in a variety of difficult combinatorial problems, such as scheduling. The reinforcement learning methods we will adopt to understand natural navigation should have direct engineering application in the development of intelligent robotic systems.