This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Efficient Risk-sensitive Planning via Entropic Risk Measures
Created by
Haebom
Author
Alexandre Marthe (ENS de Lyon, UMPA-ENSL), Samuel Bounan (UMPA-ENSL, MC2), Aur elien Garivier (UMPA-ENSL, MC2), Claire Vernade
Outline
This paper focuses on risk-sensitive planning, which seeks policies that maximize a tail-risk-sensitive metric in a Markov decision process (MDP). For widely used and interpretable metrics such as threshold probability or (conditional) value-at-risk (VaR), such optimization can be very costly. Previous studies have shown that only the entropy risk measure (EntRM) can be efficiently optimized via dynamic programming, but this requires the selection of parameters that are difficult to interpret. In this paper, we show that computing the entire set of optimal policies for EntRM over parameter values provides an accurate approximation to the metric of interest. We demonstrate that this optimality frontier can be efficiently computed thanks to a novel structural analysis and the soft nature of entropy risk. Experimental results demonstrate that the proposed approach achieves robust performance in a variety of decision scenarios.
Takeaways, Limitations
•
Takeaways:
◦
We present a novel method to efficiently approximate various indices that emphasize tail risk using entropy risk measurement (EntRM).
◦
We present a method to effectively calculate optimality fronts, enabling us to identify optimal policies for different risk preferences.
◦
Demonstrated robust performance in a variety of decision-making scenarios.
•
Limitations:
◦
A detailed analysis of the computational complexity of the proposed method may be lacking.
◦
Further validation of the generalizability of the experimental results is needed.
◦
EntRM's guidance on parameter selection may be unclear.