Learning behavior models of other agents from observations is challenging because agents typically do not act based on observ- able states alone, but usually take their internal, for external agents unobservable, states such as desires, motivations, preferences, and others into account. We propose a novel approach to on- line agent model learning that works incrementally with limited data, provides fine-grained and interpretable descriptions of the agent’s behavior, and, most importantly, is able to hypothesize agent-internal states to better explain observed behavioral trajec- tories. We show in various proof-of-concept experiments that our method avoids the pitfalls of common agent-modeling strategies when agent-internal states govern behavior and is able to build accurate and interpretable behavior models.
@inproceedings{lymperopoulosscheutz24aamas,
title={Oh, Now I See What You Want: Learning Agent Models with Internal States from Observations},
author={P. Lymperopoulos and M. Scheutz},
year={2024},
booktitle={Proceedings of AAMAS},
url={https://hrilab.tufts.edu/publications/lymperopoulosscheutz24aamas.pdf}
}