Melting Pot

Multi-agent social scenario benchmark from Google DeepMind. Tests cooperation, competition, deception, and trust with up to 16 agents.

Install:

pip install -e ".[meltingpot]"

Paradigm:

Multi-agent (simultaneous)

Stepping:

SIMULTANEOUS

Note:

Linux/macOS only

Category	Substrates
Public Goods	clean_up, commons_harvest (closed/open/partnership)
Cooperation	collaborative_cooking (7 layouts), coop_mining, chemistry (4 variants), boat_race
Coordination	bach_or_stravinsky, pure_coordination, rationalizable_coordination, stag_hunt
Competition	paintball (capture_the_flag/king_of_the_hill), territory (3 variants)
Social Dilemma	prisoners_dilemma, chicken, running_with_scissors (3 variants)
Other	allelopathic_harvest, coins, daycare, externality_mushrooms, factory_commons, fruit_market, gift_refinements, hidden_agenda, predator_prey (4 variants)

Citation

@article{leibo2021meltingpot,
  author       = {Joel Z. Leibo and Edgar Du{\'e}{\~n}ez-Guzm{\'a}n and Alexander Sasha Vezhnevets and John P. Agapiou and Peter Sunehag and Raphael Koster and Jayd Matyas and Charles Beattie and Igor Mordatch and Thore Graepel},
  title        = {Scalable Evaluation of Multi-Agent Reinforcement Learning with Melting Pot},
  journal      = {CoRR},
  volume       = {abs/2107.06857},
  year         = {2021},
}