Intention Inference Under Execution Noise: Separating Aleatoric and Epistemic Uncertainty in Social Dilemmas

In noisy social dilemmas like the Iterated Prisoner's Dilemma (IPD), agents often struggle to distinguish between an opponent's deliberate defection and a stochastic execution error. Standard Markov Decision Process (MDP) models treat all observed actions as ground truth, which leads to systematic over-retaliation. This paper asks whether explicitly modeling latent intentions as hidden states within a Partially Observable MDP (POMDP) framework can improve cooperation and strategic stability under noise.

The authors propose a POMDP solved using Active Inference (AIF). In this framework, the agent maintains a probabilistic model of the opponent's reactive policy and treats executed actions as noisy observations of latent intentions. The agent uses the expected free energy—decomposed into pragmatic value (goal-directed) and epistemic value (information-seeking)—to balance the need to learn the opponent's strategy with the desire to maximize payoffs. The transition model is learned online using Dirichlet priors, allowing the agent to adapt its beliefs about the opponent's intent over time.

Findings

The POMDP formulation provides a significant advantage over standard MDPs when playing against conditionally cooperative opponents, as it enables principled, Bayesian-informed forgiveness. However, the study identifies a critical noise threshold where mutual intention inference becomes unstable. Beyond this threshold, the agents' belief-driven dynamics lead to a correlated collapse of cooperation. Furthermore, the authors demonstrate that this representational choice is not a universal solution: against exploitable opponents, the POMDP's tendency to attribute defections to noise can be a liability, as it delays the agent's transition to an exploitative strategy.

This research highlights that the representational structure of an agent's world model fundamentally changes its social behavior. By separating aleatoric uncertainty (noise in execution) from epistemic uncertainty (lack of knowledge about intent), the authors provide a formal mechanism for forgiveness that is more adaptive than fixed-rule strategies like Tit-for-Tat. It suggests that in complex multi-agent systems, the ability to perform 'Theory of Mind'—inferring the hidden causes of observed behavior—is essential for sustaining cooperation, provided the agents can correctly identify when the environment is too noisy to support such inferences.