I'd like to propose a new Melting Pot substrate focused on cooperation around degrading shared infrastructure.
The core social dilemma would be that all agents benefit from infrastructure remaining operational, while maintenance has an individual opportunity cost. Agents can therefore maintain, free-ride, condition their contribution on others, or exploit periods of high infrastructure health.
A first version could include:
- a shared infrastructure health state that decays over time;
- maintenance actions that restore health at an individual cost;
- collective productivity/reward that depends on infrastructure health;
- heterogeneous roles or maintenance costs to create division-of-labour and fairness pressures;
- evaluation variants with unfamiliar partners and different mixtures of cooperative, conditional, and free-riding background policies.
The main research question would be whether agents generalize maintenance/cooperation strategies to unfamiliar partners and changed social compositions, rather than simply learning a fixed contribution convention.
This seems distinct from existing commons-harvest and factory substrates because the shared resource is not primarily depleted by consumption; instead, agents face a recurring costly-upkeep problem where neglect degrades the environment for everyone.
If this direction is in scope for Melting Pot, I would be happy to prototype the substrate, human-play/debug path, tests, documentation, and a small set of scripted/Puppeteer policies and evaluation scenarios. Before implementing it, I'd appreciate maintainer feedback on whether this is a useful social interaction to add and whether there are design constraints I should follow.
I'd like to propose a new Melting Pot substrate focused on cooperation around degrading shared infrastructure.
The core social dilemma would be that all agents benefit from infrastructure remaining operational, while maintenance has an individual opportunity cost. Agents can therefore maintain, free-ride, condition their contribution on others, or exploit periods of high infrastructure health.
A first version could include:
The main research question would be whether agents generalize maintenance/cooperation strategies to unfamiliar partners and changed social compositions, rather than simply learning a fixed contribution convention.
This seems distinct from existing commons-harvest and factory substrates because the shared resource is not primarily depleted by consumption; instead, agents face a recurring costly-upkeep problem where neglect degrades the environment for everyone.
If this direction is in scope for Melting Pot, I would be happy to prototype the substrate, human-play/debug path, tests, documentation, and a small set of scripted/Puppeteer policies and evaluation scenarios. Before implementing it, I'd appreciate maintainer feedback on whether this is a useful social interaction to add and whether there are design constraints I should follow.