Effects of communication by RL-agents in a single-patch foraging environment
Keywords
Loading...
Authors
Issue Date
2025-08-17
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Abstract
This thesis examines how active, cost-aware communication influences cooperation in a two agent, singlepatch
foraging environment under partial observability. A dual-channel mechanism is introduced, enabling
agents to independently control message sending (“willingness”) and reception (“attention”), with metabolic
penalties applied to both. Agents are trained using the Deep Deterministic Policy Gradient (DDPG) algorithm,
and evaluated across varying communication costs, resource abundance, and proportions of social
welfare in the reward function.
Communication declined almost linearly as costs increased, with a small unexplored performance spike at low
non-zero costs. Resource abundance produced a non-monotonic effect: in resource-scarce settings, communication
reduced performance, while in resource-rich settings it improved it, with the tipping point well above
the level at which communication first emerged. Social welfare weighting had a strong influence. Minimal
communication occurred with purely individual rewards, rising to a maximum at moderate-to-high welfare
proportions before plateauing. Overall, social welfare effects were stronger than those of direct communication
channels, indicating that reward structure can act as a strong indirect communication mechanism.
Findings suggest that the value of communication depends on environmental constraints and reward composition,
and that jointly designing both is key to fostering efficient, stable cooperation in multi-agent
reinforcement learning.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen