Researchers
Principal Investigator
Co-Principal Investigator
Student
This project is funded by the Cooperative AI Foundation (CAIF).
Project Summary
The rapid development of AI necessitates incisive evaluation of agent safety and beneficence, particularly regarding emergent properties in strategic multiagent settings. We propose to devise metrics and protocols for gauging the cooperation effectiveness (both capability and inclination to cooperate) among AI algorithms in environments featuring both collaborative and competitive elements. The main contribution will be a practical toolkit adapting the recently introduced meta-game evaluation framework for advanced AI to the statistical characterization of relevant metrics. We will demonstrate applicability of the approach to advanced AI including deep reinforcement learning and LLMs in complex domains, such as automated negotiation.