Introduction
Artificial Intelligence agents often struggle with the "unwritten rules" of human social interaction. Traditional models typically follow rigid, hand-coded rules or optimize purely for mathematical rewards, frequently ignoring the affective (emotional and social) context humans use for coordination. This research introduces a framework that balances "fast" social intuition with "slow" strategic planning by integrating Affect Control Theory (ACT) into a probabilistic decision-making model called BayesAct.
Methodology
The proposed model treats social interactions as a Partially Observable Markov Decision Process (POMDP). It leverages three core components:
- Affect Control Theory (ACT): A sociological foundation that maps identities and behaviors into a three-dimensional "EPA" space: Evaluation (good vs. bad), Potency (strong vs. weak), and Activity (active vs. inactive).
- BayesAct: This framework generalizes ACT by treating EPA values as probability distributions, allowing agents to manage uncertainty and learn during interactions.
- POMCP-C: To solve complex planning, the authors use a continuous variant of Monte-Carlo Tree Search (MCTS). This allows the agent to "look ahead" and simulate future states by dynamically discretizing continuous action spaces.
Dual-Process Reasoning: System 1 vs. System 2
The model implements a dual-process approach. It uses "Normative Action Bias"—what is socially expected—as a heuristic to guide the MCTS. When computational resources are low (low timeout), the agent relies on this "System 1" affective heuristic to make quick, socially aligned decisions. Given more time ("System 2"), the agent explores non-normative actions that may yield higher individual rewards, even if they deviate from cultural expectations.
Key Insights and Experimental Results
The framework was evaluated using the Indiana 2002-2004 ACT Lexicon and simulated social dilemmas like the Repeated Prisoner’s Dilemma and "CoRobots" (a coordination game similar to Battle of the Sexes).
Social Norms as Heuristics
Cultural prescriptions serve as powerful search heuristics rather than just constraints. They allow agents to prune massive action spaces and focus on socially sensible choices. In the Prisoner's Dilemma, the BayesAct agent successfully inferred its partner's identity (e.g., "friend" vs. "scrooge") through affective signals and adjusted its play accordingly.
The Identity Rigidity Paradox
A surprising finding in the CoRobots experiment was that agents who did not know their own identity initially coordinated better than those who did. This "Identity Rigidity Paradox" suggests that uncertainty about one's own role allows for greater flexibility in adopting complementary roles (e.g., switching between leader and follower) to achieve a common goal.
Emergent Deception and Manipulation
Deception arises naturally when an agent "fakes" an identity to influence others—a tactic known as altercasting. For instance, a "weaker" agent with high planning resources could manipulate a "stronger" agent into following its lead approximately 48% of the time, significantly increasing its individual reward.
Conclusion and Future Work
The research demonstrates that affective reasoning can be integrated into high-dimensional planning tasks, with the POMCP-C algorithm handling up to an 18-dimensional state space. Future work will focus on optimizing the implementation for real-time human-robot interaction and testing these agents against human participants to evaluate if their strategic deviations are perceived as human-like or erratic.
Summary updated