EnvCraft: A Modular Environment Design Framework for Investigating Prosocial Incentives

<p dir="ltr">Developing approaches for ensuring that the collective be- haviour of Multi-Agent Systems (MAS) is beneficial for all human stake- holders is a critical challenge for the safe deployment of MAS in real- world settings, where the goals of agent owners may partially confli...

Full description

Saved in:
Bibliographic Details
Main Author: Daniel Collins (145022) (author)
Other Authors: Conor Houghton (580148) (author), Nirav Ajmeri (10202267) (author)
Published: 2025
Subjects:
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:<p dir="ltr">Developing approaches for ensuring that the collective be- haviour of Multi-Agent Systems (MAS) is beneficial for all human stake- holders is a critical challenge for the safe deployment of MAS in real- world settings, where the goals of agent owners may partially conflict. A key direction towards addressing this challenge is understanding how to reliably incentivise prosocial behaviour among self-interested learning agents with partially conflicting goals through simulation experiments. However, methods for incentivising certain behaviours can lead to un- intended consequences if the underlying influence of environmental and social pressures on behaviour are poorly understood. Avoiding side effects of behavioural incentives is challenging in the real world, where environ- mental and social pressures are dynamic and constantly changing.</p><p dir="ltr">In this paper, we present ongoing work developing EnvCraft, a framework for composing diverse multi-agent environments from modular build- ing blocks that are parametrised to enable precise control and systematic variation of environment and population conditions. We detail re- quirements of EnvCraft for supporting robust evaluation of methods for incentivising prosocial behaviour, and for characterising the relation- ships between incentives, environment conditions and learned behaviour, through systematic exploration of the environment configuration parameters.</p>