The field of cooperative AI is concerned with advancing the cooperative intelligence of AI systems. We informally define an agent’s cooperative intelligence as its ability to achieve its goals in ways that also promote social welfare, across diverse environments and partners. The Melting Pot Contest presents a variety of settings in which, to attain a high per-capita return, agents must engage in cooperative behaviour, such as coordinating on high social welfare equilibria or sanctioning defectors.
More concretely, participants will be required to train and submit "focal" agent populations for each of four game environments from the Melting Pot suite. These focal populations will then interact with unfamiliar "background" populations of agents in the test scenarios. The contest will rank submissions based on average per-capita returns across scenarios in both "resident mode" where focal agents are the majority, and "visitor mode" where they are the minority. The environments include iterated social dilemmas and mixed-motive coordination games with up to sixteen simultaneous players.
Baseline implementations will be provided as starting points, along with tools for local debugging and visualisation. The 130+ day contest timeline includes a development phase for iterative improvement and a final test phase for official scoring. The contest is hosted via AICrowd in collaboration with colleagues from Google DeepMind and MIT, and will offer $10,000 in prizes to top performers, plus up to $50,000 in compute credits and $10,000 in travel grants to support those from underrepresented groups. Top performers will also be invited to co-author a report on the contest, to be submitted to the NeurIPS 2024 Datasets & Benchmarks track. Further details regarding the rules of the contest and information about how to enter can be found here.
The combination of mixed-motive scenarios, generalisation testing, and large populations makes Melting Pot a uniquely challenging and relevant multi-agent reinforcement learning benchmark. By hosting this contest at NeurIPS 2023, we aim to drive progress in cooperative AI and multi-agent learning, build consensus on metrics for cooperation, and engage the wider research community. Ultimately we believe successfully developing cooperative intelligence in artificial agents can lead to technologies that complement and enable fairer, flourishing societies.
In keeping with this motivation, we are committed to making this contest as accessible and inclusive as possible by providing compute resources and support for researchers from underrepresented groups in AI. Please see the contest website for details on applying for compute credits and other assistance. We aim to reduce financial and technical barriers to enable talented researchers from diverse backgrounds to participate and help the boundaries of multi-agent reinforcement learning for cooperative AI. Please contact us if you have any further questions about taking part. Good luck!
August 31: Development phase begins
September 23: Compute credits application closes
September 25: Compute funding provided to selected teams
November 10: Development phase ends
November 13: Submissions selected by participants
November 14: Evaluation phase begins
November 20: Evaluation phase ends
November 30: Winners notified
December 10-16: NeurIPS