Prediction-Aware Learning in Multi-Agent Systems
Authors: Aymeric Capitaine, Etienne Boursier, Eric Moulines, Michael I. Jordan, Alain Oliviero Durmus
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we empirically demonstrate the effectiveness of POMWU in a traffic routing experiment. |
| Researcher Affiliation | Academia | 1Centre de Mathématiques Appliquées CNRS École polytechnique Palaiseau, 91120, France 2Inria Saclay, Université Paris Saclay, LMO Orsay, 91400, France 3Inria Paris, Ecole Normale Supérieure, PSL Research University Paris, 75, France. Correspondence to: Aymeric Capitaine <EMAIL>. |
| Pseudocode | Yes | The pseudo-code of POMWU is displayed in Algorithm 1. Algorithm 1 Optimistic MWU with predicted contexts (POMWU) for agent j [J]. |
| Open Source Code | No | The network topology, the cost coefficients zp,ℓ> 0 as well as the quantities qj > 0 to be sent are downloaded from https://github.com/sessap/contextualgames/ tree/main/Sioux Falls Net. This link is for dataset/problem parameters, not the authors' source code for the methodology. No other mention of code availability is found. |
| Open Datasets | Yes | We illustrate the performances of POMWU on the Sioux Falls routing problem from Le Blanc et al. (1975) with the parameters from Sessa et al. (2019). The network topology, the cost coefficients zp,ℓ> 0 as well as the quantities qj > 0 to be sent are downloaded from https://github.com/sessap/contextualgames/ tree/main/Sioux Falls Net. |
| Dataset Splits | No | In our experiment, there are m > 0 random contexts denoted Z = {z1, . . . , zm}. At each round t [T], agents observe X0 t Rb and predict with a logistic regression ˆZt Z... The simulation is run over T = 2.104 timesteps. The paper describes a simulation environment rather than a fixed dataset with train/test/validation splits. |
| Hardware Specification | No | No specific hardware details (like GPU/CPU models or specific cloud instances) are mentioned in the paper. |
| Software Dependencies | No | The paper describes using a logistic regression and stochastic gradient descent, but does not provide specific software names or version numbers (e.g., Python, PyTorch, TensorFlow versions or other libraries). |
| Experiment Setup | Yes | The simulation is run over T = 2.104 timesteps. In our experiment, there are m = 5 states of nature. For each player j [J], we let Aj be the K = 5 shortest paths connecting nj N to mj N. we are left with J = 91 agents. It also discusses the learning rate η in Proposition 6 and 7, e.g., If all agents use Algorithm 1 with a learning rate η = (4(J 1)) 1. |