Search papers, labs, and topics across Lattice.
This paper addresses the challenge of intention inference in noisy social dilemmas by introducing a Partially Observable Markov Decision Process (POMDP) framework that distinguishes between aleatoric and epistemic uncertainty. By modeling opponent intentions as latent states and executed actions as noisy observations, the authors demonstrate that this approach can mitigate over-retaliation in scenarios like the Iterated Prisoner's Dilemma. Key findings reveal a critical noise threshold that influences cooperation dynamics, showing that intention inference can provide strategic advantages in specific contexts but may also lead to belief-driven collapse under certain conditions.
A critical noise threshold in social dilemmas reveals when intention inference can either enhance cooperation or trigger collapse, depending on the context.
In noisy social dilemmas, intended actions are stochastically corrupted before execution, so an observed defection may reflect hostile intent or action error. Standard Markov Decision Process (MDP) formulations treat executed actions as states, structurally precluding this distinction and causing systematic over-retaliation. We introduce a Partially Observable MDP (POMDP) formulation encoding opponent intentions as latent states and executed actions as noisy observations, solved within the active inference (AIF) framework with a cost function that decomposes into epistemic and pragmatic components that jointly address inferring current intent and learning how intent evolves. In the Iterated Prisoner's Dilemma with symmetric noise, we derive a critical noise threshold governing cooperation collapse, connecting it to a fixed-point condition on learned priors. Experiments reveal that the value of intention inference is context-dependent: the POMDP provides consistent advantages against conditionally cooperative opponents, but mutual intention inference under sufficient noise produces correlated belief-driven collapse. The advantage is specific to games where intent attribution is decision-relevant.
