Multi-Agents Reinforcement Learning In Iterative Voting - Archive ouverte HAL Accéder directement au contenu
Autre Publication Scientifique Année : 2019

Multi-Agents Reinforcement Learning In Iterative Voting

Résumé

In this paper we tackle the challenge of Multi Agents Reinforcement Learning (MARL) in a situation of collective social choice. We evaluate the learning performance of multiple independent learning agents interacting in an iterative plurality voting game, which is a competitive game where each agent has her private ordered preferences over a number of questions, the players submit their voting action, and the winner has the highest number of votes calculated with plurality rule. Agents receive a reward signal depending on the winner of the previous iteration. Each agent's goal is to maximize her individual long term reward. We show in our simulations that the population of agents learn to make a better collective decision. We model each voting agent as a multi-armed bandit machine and discuss the parameters that affect the outcome of the elections conducted by agents exploring under-greedy and Upper Confidence Bound (UCB) learning strategies with respect to a variant of Borda score, namely Aggregated Score Index (ASI) that reflects the overall satisfaction and the quality of the winner according to social choice criteria. We illustrate the effect of two different reward functions on the exploration strategies and discuss the impact of the size of action space and number of agents on the learning process.
Fichier principal
Vignette du fichier
Layka.MARL_for_Strategic_Voting___Thesis.pdf (874.48 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04277513 , version 1 (14-11-2023)

Identifiants

  • HAL Id : hal-04277513 , version 1

Citer

Loujain Liekah, Umberto Grandi. Multi-Agents Reinforcement Learning In Iterative Voting. 2019. ⟨hal-04277513⟩
72 Consultations
25 Téléchargements

Partager

Gmail Facebook X LinkedIn More