BayesACT: Socio-Affective Bayesian Integration

• BayesACT is a socio-affective computational framework that extends classical Affect Control Theory by embedding affective alignment in a Bayesian, decision-theoretic architecture.
• It supports adaptive decision-making under uncertainty by combining denotative (symbolic) and connotative (affective) reasoning via particle filtering and Monte Carlo Tree Search.
• BayesACT models human-like social behaviors in interactive contexts, with empirical validation in simulations and applications like tutoring and assistive technologies.

BayesACT is a socio-affective computational framework that generalizes classical Affect Control Theory (ACT) by embedding affective alignment in a Bayesian, decision-theoretic architecture. The model supports reasoning and planning under uncertainty about social identities, affective dynamics, and environmental states, explicitly integrating affective coherence with goal-driven, reward-maximizing action selection. BayesACT agents can model, infer, and adapt to human-like socio-emotional patterns in diverse interactive contexts, unifying symbolic (denotative) and affective (connotative) reasoning within partially observable Markov decision processes (POMDPs) (Hoey et al., 2013, Jung et al., 2017, Hoey et al., 2019).

1. Theoretical Foundations: ACT and Bayesian Generalization

Affect Control Theory (ACT) posits that each social entity—identity, behavior, setting—is associated with a cultural “fundamental sentiment” in Evaluation–Potency–Activity (EPA) space, with each dimension spanning roughly [−4.3, +4.3]. An Actor-Behavior-Object (A–B–O) event updates numerical EPA values for all three entities using empirically derived impression formation equations:

• Fundamental sentiment: $\mathbf{f} = (f_E, f_P, f_A) \in \mathbb{R}^3$
• Transient impression: $$\boldsymbol{\tau} = M G(\mathbf{f}) \in \mathbb{R}^9$$

The core ACT postulate is the minimization of deflection, the squared (often Mahalanobis) distance between fundamental and transient sentiments:

$D = \sum_{i=1}^9 w_i (f_i - \tau_i)^2$

BayesACT lifts all ACT states to random variables and adopts a full belief distribution, e.g., $p_t(\mathbf{f})$, over fundamentals and transients. This enables representation of uncertainty about affective identities, supports inference and adaptation, and allows flexible integration with application-specific objectives (Hoey et al., 2013, Jung et al., 2017, Hoey et al., 2019).

2. Formal Model: Joint Affective–Propositional State and Inference

BayesACT defines social interaction as a POMDP over state variables:

• Denotative state $X$: structured, symbolic content (role, task, world state)
• Connotative state $Y$: vector of EPA sentiments for all entities
• Observations: $O_t$ (denotative, e.g., actions), $E_t$ (emotional, e.g., prosody)
• Actions: $A_t^{den}$ (symbolic), $A_t^{con}$ (affective signal, EPA)

The joint transition and observation model is:

$$\boldsymbol{\tau} = M G(\mathbf{f}) \in \mathbb{R}^9$$0

where $$\boldsymbol{\tau} = M G(\mathbf{f}) \in \mathbb{R}^9$$1 is the deflection potential that regularizes joint states toward affective coherence (Hoey et al., 2019).

Belief update proceeds via Bayes’ rule, filtering for both denotative and connotative state components:

$$\boldsymbol{\tau} = M G(\mathbf{f}) \in \mathbb{R}^9$$2

Particle filtering or Gaussian mixture approximations are employed to maintain tractable inference (Hoey et al., 2013, Hoey et al., 2019).

3. Deflection, Utility, and Dual-System Action Selection

Decision-making in BayesACT agents balances external goal achievement and affective identity maintenance via a composite utility function:

$$\boldsymbol{\tau} = M G(\mathbf{f}) \in \mathbb{R}^9$$3

where $$\boldsymbol{\tau} = M G(\mathbf{f}) \in \mathbb{R}^9$$4 is application-specific (e.g., tutoring performance, game payoff), $$\boldsymbol{\tau} = M G(\mathbf{f}) \in \mathbb{R}^9$$5 is deflection, and $$\boldsymbol{\tau} = M G(\mathbf{f}) \in \mathbb{R}^9$$6 tunes the relative weight (Jung et al., 2017, Hoey et al., 2013, Hoey et al., 2019).

Action selection is performed with a softmax policy over computed $$\boldsymbol{\tau} = M G(\mathbf{f}) \in \mathbb{R}^9$$7-values:

$$\boldsymbol{\tau} = M G(\mathbf{f}) \in \mathbb{R}^9$$8

Adaptive tradeoffs between affective and denotative drives are supported by dynamically adjusting $$\boldsymbol{\tau} = M G(\mathbf{f}) \in \mathbb{R}^9$$9 based on belief entropy: system “leans” toward affective (System 1) or propositional (System 2) control as a function of contextual uncertainty (Hoey et al., 2019). A representative rule is

$D = \sum_{i=1}^9 w_i (f_i - \tau_i)^2$0

where $D = \sum_{i=1}^9 w_i (f_i - \tau_i)^2$1 and $D = \sum_{i=1}^9 w_i (f_i - \tau_i)^2$2 are the entropies of denotative and connotative beliefs, respectively.

4. Inference and Planning Algorithms

The standard BayesACT loop follows:

1. Observe new action/affect signals $D = \sum_{i=1}^9 w_i (f_i - \tau_i)^2$3.
2. Update belief $D = \sum_{i=1}^9 w_i (f_i - \tau_i)^2$4 via observation likelihood and particle-based filtering.
3. Compute entropies $D = \sum_{i=1}^9 w_i (f_i - \tau_i)^2$5, $D = \sum_{i=1}^9 w_i (f_i - \tau_i)^2$6; adapt $D = \sum_{i=1}^9 w_i (f_i - \tau_i)^2$7 accordingly.
4. For each candidate $D = \sum_{i=1}^9 w_i (f_i - \tau_i)^2$8, evaluate $D = \sum_{i=1}^9 w_i (f_i - \tau_i)^2$9.
5. Sample/select optimal action according to $p_t(\mathbf{f})$0.
6. Act and repeat (Hoey et al., 2013, Jung et al., 2017, Hoey et al., 2019).

In problem domains with continuous affective state, planning is handled via Monte Carlo Tree Search adapted to the augmented belief space and reward structure. Expansion and sampling are biased toward trajectories predicted to incur low deflection.

5. Empirical Validation and Behavioral Phenomena

BayesACT has demonstrated the ability to replicate diverse “human-like” behaviors in simulated social dilemmas and assistive contexts:

• Iterated Networked Prisoner’s Dilemma (INPD): BayesACT agents display four of five key empirical patterns observed in human INPD play—network invariance, anti-correlation of cooperation and reward, stratified player types, and moody conditional cooperation. Network invariance is validated by G-test $p_t(\mathbf{f})$1 in $p_t(\mathbf{f})$2 of cases, and stratification by matching the “mixed > mostly > pure” agent type ordering across all settings. However, BayesACT does not reproduce the empirical decline in cooperation over time (Jung et al., 2017).
• Identity learning and shape-shifting: Bayesian belief updates enable agents to infer and track dynamic identities, adapting actions and affect accordingly (Hoey et al., 2013).
• Assistive and tutoring agents: Empirical studies in exam-practice tutoring and cognitive assistance (e.g., prompting persons with dementia) show that BayesACT policies yield lower deflection and superior affective alignment, with measurable improvements in user-experience metrics over fixed-strategy baselines (Hoey et al., 2013).
• Cognitive biases: BayesACT, when equipped with appropriate connotative priors, reproduces core psychological effects (fairness under uncertainty, cognitive dissonance, and conformity dynamics) as Bayesian inference over affective-coherent beliefs. These findings extend to the unification of exploration strategies in RL: random exploration aligns with high denotative uncertainty, while intrinsic reward shaping can be interpreted as minimizing expected deflection (Hoey et al., 2019).

6. Applications and Extensions

Research demonstrates BayesACT’s application as a domain-general affective “plug-in” for interactive agents:

• Social dilemmas: Explaining deviation from pure payoff maximization, modeling group norms and reciprocity (Jung et al., 2017).
• Human-assistive technologies: Guiding prompting and dialogue generation to maintain user engagement, self-efficacy, and affective comfort (Hoey et al., 2013).
• RL and AI safety: Shaping agent policy spaces via affective constraints to encourage socially normed exploration and adaptive behaviors, with interpretability rooted in sociological theory (Hoey et al., 2019).

Ongoing research focuses on richer models of identity dynamics, maintaining separate affective beliefs for multiple partners, expanding to nontrivial social networks, and integrating temporal asynchrony and multi-modal affective observation.

7. Open Issues and Future Research

Empirical gaps remain—specifically, the inability to reliably produce declining cooperation rates over repeated social dilemma rounds. Enhancements such as accelerated identity updating or individualized EPA beliefs for each neighbor are proposed to address these phenomena (Jung et al., 2017).

Future directions include:

• Generalization to large-scale, heterogeneous social networks.
• Real-world deployment in negotiation, social robotics, and online collaboration.
• Bayesian learning of impression-formation coefficients and sentiment dictionaries, replacing fixed empirically derived components with fully adaptive models (Hoey et al., 2013, Hoey et al., 2019).

BayesACT establishes a principled, quantitatively validated template for affective-cognitive integration in artificial agents, replicating critical facets of human social behavior and supporting robust, culturally informed human–AI interaction.