Notice: 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. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
On the Convergence of Single-Timescale Actor-Critic
Authors: Navdeep Kumar, Priyank Agrawal, Giorgia Ramponi, Kfir Y. Levy, Shie Mannor
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Theoretical | We analyze the global convergence of the single-timescale actor-critic (AC) algorithm for the infinite-horizon discounted Markov Decision Processes (MDPs) with finite state spaces. To this end, we introduce an elegant analytical framework for handling complex, coupled recursions inherent in the algorithm. Leveraging this framework, we establish that the algorithm converges to an ϵ-close globally optimal policy with a sample complexity of O(ϵ 3). |
| Researcher Affiliation | Academia | Navdeep Kumar Technion Priyank Agrawal, Columbia University Giorgia Ramponi University of Zurich Kfir Levy Technion Shie Mannor Technion |
| Pseudocode | Yes | Algorithm 1 Single Time Scale Actor Critic Algorithm Input: Stepsizes ηk, βk 1: while not converged; k = k + 1 do 2: Sample s dπθk , a πθk( |s) and get the next state-action s P( |s, a), a πθk( |s ) . 3: Policy update: θk+1(s, a) = θk(s, a) + ηk(1 γ) 1A(s, a), where A(s, a) = Q(s, a) v(s) and v(s) = P a πθk(a|s)Q(s, a). 4: Q-value update: Q(s, a) = Q(s, a) + βk h R(s, a) + γQ(s , a ) Q(s, a) i . 5: end while |
| Open Source Code | Yes | All MDPs have randomly generated transition kernel and reward function, with codes available at https://anonymous.4open.science/r/AC-C43E/. |
| Open Datasets | No | All MDPs have randomly generated transition kernel and reward function, with codes available at https://anonymous.4open.science/r/AC-C43E/. For simplicity, the samples are generated uniformly instead of discounted occupation measure. |
| Dataset Splits | No | The paper mentions generating random MDPs and sampling uniformly, but does not specify any dataset splits like train/test/validation percentages or counts. |
| Hardware Specification | No | The NeurIPS checklist states that the paper is purely theoretical in nature and thus does not include experiments (Answer: NA for compute resources), implying no specific hardware was used for experimental runs. |
| Software Dependencies | No | The paper mentions that codes are available but does not list specific software dependencies with version numbers (e.g., Python 3.8, PyTorch 1.9) required to reproduce the numerical simulations. |
| Experiment Setup | Yes | Figure 2: Convergence Rate of Algorithm 1, on random MDP with state space =50, action space = 5, learning rate ηk = βk = k a; Figure 3: Convergence Rate of Algorithm 1, on random MDP with state space =5, action space = 2, learning rate 10ηk = βk = k a; Figure 4: Convergence Rate of Algorithm 1, on random MDP with state space =20, action space = 5, learning rate ηk = βk = k a. |