SNAN Spiking Neuron-Astrocyte Grid Navigation Model

Paper context

Paper title: Dual-Timescale Memory in a Spiking Neuron-Astrocyte Network for Efficient Navigation Abstract: Biological agents navigate complex environments by combining long-term memory of successful actions with short-term suppression of recently visited locations-a capability that remains difficult to replicate in artificial systems, especially under partial observability. Inspired by the complementary timescales of neural and astrocytic dynamics, we introduce a spiking neuron-astrocyte network (SNAN) where spike-timing-dependent plasticity (STDP) reinforces successful action sequences on a distant time scale, while astrocytic calcium transients suppress recently visited states on a short-term time scale, effectively blocking locations already explored. This dual-timescale memory mechanism biases the agent toward unexplored regions, accelerating goal finding without requiring explicit global statistics. We show that in grid-world navigation tasks with extreme partial observability, SNAN reduces median path length by up to sixfold and drastically improves goal completion rates compared to baseline agents. The astrocytic modulation inherently mitigates the exploration-exploitation trade-off as an emergent consequence of local state suppression. This kind of local sensory data modulation can be considered as a new type of working memory referred to as a "Topological-Context Memory". To validate hardware feasibility using neuromorphic approaches, we map STDP to a memristive VTEAM model and implement a subset of the network on a crossbar array, achieving order-of-magnitude gains in sp Passages referencing this figure: outline the limitations of the proposed method, and provide insights for future work. 2 Methods 2.1 Problem Statement We consider the problem of navigating an agent in a partially observable grid environment, where the agent lacks complete, real-time access to the full state of its surroundings, except for an observation that consists of a single cell of a grid environment denoting its position (Fig. 1 A). Each state can be defined by the agent’s position; thus, state space S S contains all possible agent positions. The environment’s transition function is deterministic. The action space consists of four actions A A that move the agent to each adjacent grid cell: up, down, left, and right. The agent receives no observations about the environment. However, the agent is unable to move beyon init} , and goal state, S G S_{G} are chosen randomly, with S i ​ n ​ i ​ t ≠ S G S_{init}\neq S_{G} . The trial ends when the agent reaches the goal state or when the time limit is reached. As a metric of the agent’s performance during a trial, we use the number of steps required for the agent to reach the goal. For each agent and environment configuration, we perform several independent trials. Figure 1: The considered grid environment and the proposed bioinspired short-term memory model implemented by a spiking neuron–astrocyte network (SNAN) for an agent navigating. (A) An agent navigating a discrete 2D grid from an initial position (green cell) to a goal (yellow cell). Arrows indicate possible movement paths along the cardinal directions. (B) SNAN architecture. (C) Schematic of the SN (H) Membrane potential of output motor neurons #1,4,6,9 synaptically connected to sensory input neuron #5. 2.2 Spiking Neuron-Astrocyte Network Model We propose a bioinspired short-term memory model