BayesAct: Affective-Cognitive Agent Model

• BayesAct is a computational framework that integrates probabilistic affective sentiment modeling with cognitive decision-making to minimize deflection.
• It employs a dual-process architecture combining fast affective (connotative) and deliberative (denotative) processes via the somatic transform for adaptive social interactions.
• The model has proven effective in tutoring, assistive technologies, and social simulations, ensuring culturally congruent agent behavior.

The BayesAct model is a computational framework for modeling affectively intelligent agents that systematically integrate affective and cognitive reasoning. It is rooted in Affect Control Theory (ACT), extended with probabilistic representations and decision-theoretic planning to enable artificial agents to maintain, infer, and adapt culturally shared affective sentiments in dynamic, uncertain environments. BayesAct incorporates dual-process models, synthesizes affective alignment with explicit decision objectives, and supports practical applications in tutoring, assistive technology, and social simulation.

1. Theoretical Foundations and Core Principles

BayesAct generalizes classical Affect Control Theory, which posits that humans strive to minimize discrepancies (“deflection”) between culturally shared, fundamental affective sentiments (identities, behaviors, objects) and transient sentiments generated by ongoing interactions. These sentiments are organized in a three-dimensional space of Evaluation (E), Potency (P), and Activity (A), collectively called EPA space.

Unlike ACT, where sentiments are single points, BayesAct represents sentiments as probability distributions over EPA latent variables. This approach enables agents to encode, update, and reason about uncertainty over identities and feelings. BayesAct formalizes deflection as

$$\delta = (s_\text{trans} - s_\text{fundamental})^T (s_\text{trans} - s_\text{fundamental}),$$

where $s_\text{trans}$ is the transient sentiment vector and $s_\text{fundamental}$ is the fundamental sentiment vector.

BayesAct agents jointly optimize for minimal expected deflection (social-psychological congruence) and explicit task or reward-based objectives, embedding a dual motivation structure.

2. Dual-Process Architecture and Somatic Transform

BayesAct operationalizes dual-process theories of agency, integrating a “System 2”-like deliberative process (reasoning over high-dimensional, denotative states) and a “System 1”-like affective process (fast, low-dimensional, connotative states). The agent’s denotative state encodes explicit symbolic information (e.g., roles, statuses), while its connotative state encodes EPA-based affect.

A central mechanism is the “somatic transform,” a potential function that bridges denotative and connotative representations. The mapping is formalized with an energy function:

$$G(x, y) = c \cdot \exp\left[-\frac{(y - M(x))^2}{\gamma^2}\right],$$

where $x$ is the denotative state, $y$ is the connotative (EPA) value, $M(x)$ is the mean cultural affect for $x$, $\gamma$ modulates uncertainty in the mapping, and $c$ is a normalizer. This enables bidirectional updating: denotative cues inform affective expectations, and observed emotions feed back to reinterpret symbolic labels.

The equilibrium between affective (somatic potential–driven) and cognitive (POMDP reward–driven) reasoning is dynamically modulated by environmental uncertainty: under high uncertainty or low denotative validity, heuristic affect-driven processing dominates.

3. Decision-Theoretic Planning and Deflection Minimization

BayesAct frames agent decision making as a Partially Observable Markov Decision Process (POMDP). Agents maintain belief distributions over affective and denotative states, updating them as new observations accrue. Action selection is driven by a combination of deflection minimization and explicit utility maximization.

Planning is implemented via Monte Carlo Tree Search, specifically the Partially Observable Monte Carlo Planning (POMCP) algorithm, which prioritizes action sequences that minimize expected deflection. The optimal policy thus inherently favors actions that simultaneously achieve task rewards and are congruent with culturally appropriate affect.

Empirical findings demonstrate that BayesAct effectively adapts to uncertain or shifting affective identities, with the model rapidly reducing identity deflection (typically within 50–200 samples), and robustly handling moderate environmental noise (noise parameter $e_0 < 1.0$). The agent’s belief converges swiftly even as user affect changes over time, exhibiting essential dynamic tracking properties.

4. Emergent Social Behavior and Simulation Results

Simulations with BayesAct in the Iterated Networked Prisoner's Dilemma (INPD) exhibit four out of five canonical properties of human play:

• Network invariance: Cooperation rates remain statistically invariant across different network topologies (confirmed via G-tests with $p > 0.05$).
• Anti-correlation of cooperation and reward: Agents that cooperate more earn lower rewards, as in human datasets (statistically significant negative Pearson coefficients).
• Player type stratification: BayesAct produces player populations distributed into the established categories (pure/mixed/mostly cooperators/defectors), matching empirical human stratification.
• Moody Conditional Cooperation (MCC): About two-thirds of simulated environments yield the characteristic hysteresis and conditionality observed in human cooperation. In contrast, imitation-based agents only reliably replicate MCC, failing in network invariance and player stratification.

The table below summarizes emergent properties:

Property	BayesAct Agents	Imitation Agents
Network invariance	Achieved	Not achieved
Coop–Reward anti-correlation	Achieved	Not achieved
Player stratification	Achieved	Seldom
Moody Conditional Cooperation	2/3 scenarios	Always

BayesAct’s integration of affective and cognitive reasoning underlies its performance across diverse social settings.

5. Applications in Human-Interactive Systems

BayesAct provides an affective “plug-in” for intelligent interactive systems:

• Intelligent Tutoring: BayesAct-driven tutors combine content-level action (“what” to teach) with affectively tuned delivery (“how” to teach, as EPA vectors). In a pilot paper, users rated BayesAct tutors as smoother and more natural, with reduced deflection relative to random-affect tutors. This corresponded with improved communication and lower affective friction.
• Assistive Technology: In cognitive assistance (e.g., COACH handwashing assistant), BayesAct adaptively selects affective prompts based on client identity and incoming feedback, outperforming fixed-action policies. When clients exhibited non-normative or extreme identities, adaptive BayesAct policies led to lower deflection and more efficient task completion.
• Socio-Technical Systems: In scenarios such as online collaboration networks, BayesAct agents can mediate interactions by aligning with emergent social norms and mitigating conflict through affectively congruent responses.
• Reinforcement Learning and Exploration: BayesAct implicitly unifies multiple exploration strategies, modulating exploration/exploitation tradeoff as a function of uncertainty in connotative and denotative states. The somatic transform introduces an “affective bonus,” steering the agent toward socially normative or information-rich actions under ambiguity.

6. Modeling Social-Psychological Phenomena and Cognitive Biases

BayesAct accounts for a range of social-psychological phenomena by explicitly modeling the probabilistic coupling between denotative state and connotative affect:

• Fairness: Under uncertainty, agents default to affective norms, yielding more equitable choices—a pattern observed in human fairness experiments.
• Cognitive dissonance: When denotative outcomes and affective self-concepts conflict, the somatic transform drives reinterpretation of symbolic states to restore emotional coherence.
• Conformity: Repeated observation of group behavior shifts denotative beliefs by connotative association, capturing phenomena such as Asch’s conformity effects. This probabilistic fusion supports context-sensitive, norm-aligned adaptation, explicating cognitive biases as emergent features of dual-process affect-control.

7. Implications and Outlook

BayesAct bridges sociological affect theory with computational models for artificial agency, enabling agents to reason and act in ways that are simultaneously goal-oriented and socially congruent. The model’s success in replicating human-like emergent properties in networked social dilemmas, alongside practical deployments in tutoring and assistive technologies, demonstrates its versatility.

By fusing probabilistic affective tracking, dynamic policy computation, and dual-process control, BayesAct provides a robust foundation for constructing agents that navigate the affective dimensions of real-world interaction. Its capacity to account for uncertainty, model dynamic identity change, and unify reinforcement learning exploration further positions BayesAct as a foundational tool for designing socio-affective agents in human–machine societies.

A plausible implication is that as artificial agents proliferate in social and collaborative contexts, BayesAct’s principles and mechanisms will be increasingly vital for ensuring seamless, trust-enhancing, and psychologically safe interaction between humans and technological agents (Hoey et al., 2013, Jung et al., 2017, Hoey et al., 2019).