Risk-Aware Policies Improve Agent Survival Under Noisy Perception

Biological systems constantly make decisions under noisy and uncertain perceptual input, where errors can have severe consequences. This research explores how artificial agents can navigate similar challenges by developing risk-aware decision policies. The study uses an Artificial Life predator-prey model, simulating foraging behavior under conditions of noisy perception.Through controlled experiments, the researchers compared agent performance under various policies, specifically focusing on how agents account for their noisy predictions. The findings clearly show that agents who blindly trust their perceptual labels experience catastrophic failures as noise levels increase. In stark contrast, agents employing strategies that explicitly consider uncertainty demonstrate significantly improved survival rates and a drastic reduction in fatal errors.The study also observed qualitative shifts in agent behavior: as uncertainty grew, agents transitioned from exploratory foraging strategies to more conservative ones. These results underscore the critical importance of uncertainty-aware decision-making for robustness in environments with unreliable perception. The model provides an interpretable analogue to robust learning with noisy labels, linking risk-sensitive foraging, ecological information use, and Artificial Life research.