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.
Pessimistic Iterative Planning with RNNs for Robust POMDPs
Created by
Haebom
Author
Maris FL Galesloot, Marnix Suilen, Thiago D. Sim ao, Steven Carr, Matthijs TJ Spaan, Ufuk Topcu, Nils Jansen
Outline
This paper proposes Pessimistic Iterative Planning (PIP), a novel framework for robust partially observable Markov decision processes (POMDPs) that account for model uncertainty. PIP computes a robust policy that considers worst-case probabilistic instances using uncertainty sets about transition and observation functions. PIP iteratively selects the worst-case probabilistic instance and computes its finite-state controller (FSC). In this paper, we propose the rFSCNet algorithm, which optimizes recurrent neural networks to compute the FSC. Experimental results demonstrate that rFSCNet outperforms existing methods in computing robust policies.
Takeaways, Limitations
•
Takeaways:
◦
An effective solution to the robust POMDP problem considering model uncertainty.
◦
The rFSCNet algorithm enables robust policy computation that outperforms existing methods.
◦
Efficiently learning finite-state controllers using recurrent neural networks.
•
Limitations:
◦
Lack of a clear analysis of the computational complexity of the proposed method.
◦
Further research is needed on the generalizability to different types of uncertainty sets.
◦
The scope of experimental evaluation may be limited.
