Learning Utilities from Demonstrations in Markov Decision Processes

Filippo Lazzati, Alberto Maria Metelli
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:32704-32770, 2025.

Abstract

Although it is well-known that humans commonly engage in risk-sensitive behaviors in the presence of stochasticity, most Inverse Reinforcement Learning (IRL) models assume a risk-neutral agent. As such, beyond $(i)$ introducing model misspecification, $(ii)$ they do not permit direct inference of the risk attitude of the observed agent, which can be useful in many applications. In this paper, we propose a novel model of behavior to cope with these issues. By allowing for risk sensitivity, our model alleviates $(i)$, and by explicitly representing risk attitudes through (learnable) utility functions, it solves $(ii)$. Then, we characterize the partial identifiability of an agent’s utility under the new model and note that demonstrations from multiple environments mitigate the problem. We devise two provably-efficient algorithms for learning utilities in a finite-data regime, and we conclude with some proof-of-concept experiments to validate both our model and our algorithms.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-lazzati25a, title = {Learning Utilities from Demonstrations in {M}arkov Decision Processes}, author = {Lazzati, Filippo and Metelli, Alberto Maria}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {32704--32770}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/lazzati25a/lazzati25a.pdf}, url = {https://proceedings.mlr.press/v267/lazzati25a.html}, abstract = {Although it is well-known that humans commonly engage in risk-sensitive behaviors in the presence of stochasticity, most Inverse Reinforcement Learning (IRL) models assume a risk-neutral agent. As such, beyond $(i)$ introducing model misspecification, $(ii)$ they do not permit direct inference of the risk attitude of the observed agent, which can be useful in many applications. In this paper, we propose a novel model of behavior to cope with these issues. By allowing for risk sensitivity, our model alleviates $(i)$, and by explicitly representing risk attitudes through (learnable) utility functions, it solves $(ii)$. Then, we characterize the partial identifiability of an agent’s utility under the new model and note that demonstrations from multiple environments mitigate the problem. We devise two provably-efficient algorithms for learning utilities in a finite-data regime, and we conclude with some proof-of-concept experiments to validate both our model and our algorithms.} }
Endnote
%0 Conference Paper %T Learning Utilities from Demonstrations in Markov Decision Processes %A Filippo Lazzati %A Alberto Maria Metelli %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-lazzati25a %I PMLR %P 32704--32770 %U https://proceedings.mlr.press/v267/lazzati25a.html %V 267 %X Although it is well-known that humans commonly engage in risk-sensitive behaviors in the presence of stochasticity, most Inverse Reinforcement Learning (IRL) models assume a risk-neutral agent. As such, beyond $(i)$ introducing model misspecification, $(ii)$ they do not permit direct inference of the risk attitude of the observed agent, which can be useful in many applications. In this paper, we propose a novel model of behavior to cope with these issues. By allowing for risk sensitivity, our model alleviates $(i)$, and by explicitly representing risk attitudes through (learnable) utility functions, it solves $(ii)$. Then, we characterize the partial identifiability of an agent’s utility under the new model and note that demonstrations from multiple environments mitigate the problem. We devise two provably-efficient algorithms for learning utilities in a finite-data regime, and we conclude with some proof-of-concept experiments to validate both our model and our algorithms.
APA
Lazzati, F. & Metelli, A.M.. (2025). Learning Utilities from Demonstrations in Markov Decision Processes. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:32704-32770 Available from https://proceedings.mlr.press/v267/lazzati25a.html.

Related Material