We investigated the hypothesis that even seemingly unrelated sources of data may contribute actionable intelligence to an agent in a world for which it has never trained. Using a virtual agent navigating a maze, we show that the patterns obtained from a mathematical object---a Halley fractal---and human generated art are better able to guide maze exploration than patterns with less structure. Further, we show that replacing half of the environmental information in the input layer of a deep Q-learning agent with information about an arbitrary fractal increases exploration ability. Overall we demonstrate the utility of combining sensory information with static internal patterns that can be obtained from different sources for generating interesting, more complex behavior.
@inproceedings{ertleetal25icdl,
title={Free Lunch? Low-Cost Intelligence Through Pattern-Guided Exploration},
author={Emily Ertle and Michael Levin and Matthias Scheutz},
year={2025},
booktitle={International Conference on Development and Learning},
url={https://hrilab.tufts.edu/publications/ertleetal25icdl.pdf}
}