General-sum games sit between fully cooperative and purely adversarial settings, and they pose a distinct challenge: agents may be tempted to deceive or exploit one another. According to a new arXiv paper, achieving high social welfare in such games requires sustained cooperation, yet a single defection can unravel that outcome.
The paper introduces reciprocal Q-augmentation as a way to support cooperation without relying on shared rewards. The abstract frames this as a departure from settings where agents are explicitly aligned, suggesting the method modifies Q-values in a reciprocal manner to encourage mutually beneficial behavior.
Because the abstract is brief, the exact mechanics and experimental results are not detailed here. The significance lies in the direction: enabling cooperative outcomes in general-sum games where agents do not have a common reward signal, which could broaden the applicability of multi-agent reinforcement learning.