Achieving mouse-level strategic evasion performance using real-time computational planning

German Espinosa, Gabrielle E. Wink, Alexander T. Lai, Daniel A. Dombeck, Malcolm A. MacIverarXiv 2022 · DOI · arXiv

Abstract

Real-time planning in dynamic, cluttered environments presents a significant challenge for artificial intelligence due to the "curse of dimensionality." While standard algorithms like Partially Observable Monte-Carlo Planning (POMCP) require massive computational resources to evaluate future states, biological systems perform strategic evasion with remarkable efficiency. This paper introduces TLPPO (Trajectories over Locations where Planning Pays Off), a biologically-inspired planning algorithm that uses environmental geometry to prune search spaces. By focusing computation only on "high-value" transition points, TLPPO achieves mouse-level survival rates against a robotic predator using three orders of magnitude less computational effort than standard AI planners.

Introduction

In a dynamic environment—such as a mouse dodging a predator in a cluttered forest—the number of possible future states grows exponentially with every step. For an agent to survive, it must make strategic decisions within millisecond-scale reaction times. Standard AI frameworks often fail this real-time constraint; for instance, POMCP may require hundreds of thousands of simulations to find a "safe" path, a budget that physical hardware cannot support in live-action scenarios. This research bridges the "complexity gap" by identifying how biological systems might use the structure of their environment to gate expensive cognitive processes like planning.

Methods: The TLPPO Algorithm

The core innovation of TLPPO is the use of environmental geometry to decide when and where to plan. The algorithm employs a graph theory metric called eigenvector centrality (eigencentrality) to measure the "connectedness" of a space.

Planning is most valuable at transitions, specifically when moving from "closed" spaces (clutter or cover) to "open" spaces. TLPPO calculates the gradient of eigencentrality across a map to identify these "points of interest." Unlike standard planners that expand a decision tree at every possible action, TLPPO only triggers tree expansion when the agent reaches one of these high-value locations. Between these points, the agent follows the shortest path, drastically reducing the number of branches sampled.

The experimental setup involves a robotic predator and a mouse in a hexagonal arena with obstacles, comparing real-world prey behavior to computational TLPPO agents.

Experimental Setup

The researchers compared TLPPO against live lab mice (Mus musculus) and a POMCP baseline in a hexagonal arena filled with obstacles. A reactive robotic predator was programmed to deliver aversive air puffs upon "capture." To establish a gold standard, 230 trajectories from live mice were recorded, showing a baseline survival rate of 86%. The computational agents (TLPPO and POMCP) were then tested in both simulation and 100 real-world experimental episodes.

Key Results

The results demonstrate a massive disparity in computational efficiency between biological-inspired planning and standard AI. To match the 86% survival rate of a live mouse, the standard POMCP algorithm required sampling approximately 175,000 branches per iteration. In contrast, TLPPO achieved a comparable 82% survival rate with only ~100 branches.

Performance metrics show TLPPO matching mouse-level survival with 100 branches, while POMCP requires 175,000. TLPPO fits within the real-time planning budget (~150ms).

When POMCP was restricted to the same real-time "thinking budget" as TLPPO (limited to 1,000 branches to fit within a ~150ms window), its survival rate plummeted to just 13%. This suggests that strategic pruning based on environmental geometry is not just an optimization, but a necessity for real-time survival in complex spaces.

Discussion and Future Work

TLPPO supports the neuroscientific theory that the brain—specifically the hippocampus—does not just store a static map but uses the structure of that map to trigger planning. By ignoring thousands of irrelevant "futures" and focusing only on locations where the environment's "openness" changes significantly, agents can achieve high-level strategic behavior on limited hardware.

Future work will address the limitations of static environments, as the current model pre-calculates points of interest based on a fixed map. Adapting TLPPO to fluid environments where obstacles move would require updating eigencentrality maps on the fly. Additionally, further research is needed to confirm whether animal brains explicitly utilize eigencentrality gradients as a trigger for cognitive planning.

Summary updated

BibTeX
@misc{espinosa2022achieving,
  author = {German Espinosa and Gabrielle E. Wink and Alexander T. Lai and Daniel A. Dombeck and Malcolm A. MacIver},
  title = {{Achieving mouse-level strategic evasion performance using real-time computational planning}},
  year = {2022},
  eprint = {2211.02700},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2211.02700},
}
Figure