Notice: This website is continuously updated as new conference proceedings and journal volumes are published. Consequently, the figures presented here may differ from those reported in our publications.
A brief overview of our methodology is provided on the Methods page, along with references to publications documenting our approach and links to the code and data used in each study.
The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates.
Spectral Learning for Infinite-Horizon Average-Reward POMDPs
Authors: Alessio Russo, Alberto Maria Metelli, Marcello Restelli
NeurIPS 2025 | Venue PDF | LLM Run Details | Input Tokens: 43,300 Total number of tokens sent to the LLM as input for this paper's analysis. | Output Tokens: 3,223 Total number of tokens produced by the LLM (including reasoning/thinking tokens) for this paper's analysis.
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we present numerical simulations that validate the theoretical analysis of both the proposed estimation procedure and the Mixed Spectral UCRL algorithm. |
| Researcher Affiliation | Academia | Alessio Russo DEIB, Politecnico di Milano EMAIL Alberto Maria Metelli DEIB, Politecnico di Milano EMAIL Marcello Restelli DEIB, Politecnico di Milano EMAIL |
| Pseudocode | Yes | Algorithm 1 Mixed Spectral Estimation. 1: Input: Trajectory set Γ {τl}L 1l 0 where for each l we have τl {pol j, al j}Nl 1j 0 2: Output: Estimated Observation model p O and Transition model {p Ta}a PA 3: for a ∈ A do |
| Open Source Code | Yes | The codebase can be found at https://github.com/alesnow97/Spectral_Learning_POMDP.git. |
| Open Datasets | No | For the generation of the different POMDPs, we adopted a similar approach to the one followed in [25]. The matrices of both the observation and transition models are randomly generated, and successive modifications are applied: |
| Dataset Splits | No | The simulation splits the interaction horizon into 10 episodes of equal length, and for each episode, we use a different belief-based policy for data collection. |
| Hardware Specification | Yes | The simulations illustrated in this work have been run on an 88 Intel(R) Xeon(R) CPU E7-8880 v4 @ 2.20GHz CPUs with 94 GB of RAM. |
| Software Dependencies | No | The paper mentions the codebase is available at a GitHub link, implying standard software like Python would be used, but it does not specify any particular software, libraries, or frameworks with their version numbers. |
| Experiment Setup | Yes | For the experiments on the regret, we adopted the following hyperparameters for the different algorithms. Mixed Spectral UCRL: length of initial episode T0 3 · 105; SM-UCRL: length of initial episode T0 3 · 105, minimum action probability ι 0.02; SEEU: length of exploration phase τ1 105, length of initial exploitation phase τ2 3 · 105. At each new episode l, the length of the exploitation phase is computed as √l · τ2, as defined in the original work. ... For the considered simulations, we adopted a discretization step size of 0.04. |