Bayesian Affect Control Theory (BayesAct)

• Bayesian Affect Control Theory (BayesAct) is a probabilistic framework that defines and updates affective states using Bayesian inference and classical affect control principles.
• It models social interactions as a POMDP, balancing affective deflection minimization with task-oriented utility through dynamic belief updates and action selection.
• Empirical applications in social dilemmas and human-computer interaction demonstrate that BayesAct improves affect alignment and enhances user satisfaction.

Bayesian Affect Control Theory (BayesAct) is a probabilistic, decision-theoretic generalization of classical Affect Control Theory (ACT) that enables artificial agents to reason about affective states, maintain and update multiple affective hypotheses, and choose actions that blend affective coherence with instrumental objectives. BayesAct formalizes socio-affective reasoning as a partially observable Markov decision process (POMDP), integrating the sociological insight that affective sentiment alignment governs much of human social interaction with the computational benefits of Bayesian inference and utility maximization. It is rooted in decades of empirical social psychology and extended to provide practical, tractable algorithms for affectively intelligent human-computer interaction, multi-agent social simulation, and applications such as tutoring and assistive devices (Hoey et al., 2013, Hoey et al., 2019, Jung et al., 2017).

1. Theoretical Foundations: From ACT to BayesAct

Classical Affect Control Theory posits that individuals act to minimize “deflection,” the squared distance between fundamental sentiments (culturally shared, identity-linked EPA vectors—Evaluation, Potency, Activity—measured empirically) and transient impressions induced by events. Each event is an (Actor, Behavior, Object) triple, and impression formation is captured by empirically derived polynomial equations:

$\tau = M \cdot G(f)$

where $f$ concatenates the EPA vectors of actor, behavior, and object, $G$ is a feature function (including cross-terms), and $M$ is a regression matrix fitted to survey data. Deflection is

$$D(f, \tau) = (f - \tau)^\top \Sigma^{-1} (f - \tau)$$

where $\Sigma$ is an empirically derived covariance (Hoey et al., 2013, Jung et al., 2017).

BayesAct reconceptualizes all affective variables (sentiments, impressions) as random variables with explicit probabilistic structure. This extension supports:

• Priors over unknown identities or affective roles, $p(f)$,
• Probabilistic observation models, $p(o\,|\,\tau)$,
• Dynamic posterior updates over sentiments and impressions,
• Action selection via expected utility combining affective and task components (Hoey et al., 2013, Hoey et al., 2019, Jung et al., 2017).

2. Formal Specification and Bayesian Inference

The BayesAct model factors the social-interaction POMDP at each step into:

Latent State

$S_t = (F_t, T_t, X_t)$

• $F_t \in \mathbb{R}^9$: Fundamental (identity) EPA vectors (Actor, Behavior, Object)
• $T_t \in \mathbb{R}^9$: Transient impressions
• $X_t$: Application-specific task state

Transitions

The state transition decomposes as:

$$P(S_{t+1} \mid S_t, B_t) = P(X_{t+1} \mid X_t, F_{t+1}, T_{t+1}, A_t) \, P(T_{t+1} \mid T_t, F_{t+1}, X_t) \, P(F_{t+1} \mid F_t, T_t, X_t, B^a_t)$$

with:

• Impression-formation: $T_{t+1}$ determined by deterministic or noisy ACT polynomial (via $M$, $G(f)$, possibly with Gaussian noise).
• Fundamental evolution: $F_{t+1}$ is drawn from a product of a deflection-minimizing potential and an inertia (slow drift) prior.
• Task evolution: $X_{t+1}$ transitions depending on task and affective variables (Hoey et al., 2013).

Observation Model

Observations $O_t = (\omega^f_t,\ \omega^x_t)$:

$$P(O_t \mid S_t, B_t) = P(\omega^x_t \mid X_t)\, P(\omega^f_t \mid F_t)$$

with observation noise modeled as Gaussian (Hoey et al., 2013).

Inference

Posterior beliefs $b_t(S)$ are updated by Bayes’ rule, typically using bootstrap particle filters for continuous ($F, T$) and discrete ($X$) states. Posterior over $F'$, conditioned on impression and prior, is Gaussian with

$$\Sigma' = (E^{-1} + \Sigma_f^{-1})^{-1}, \quad \mu' = \Sigma' \left(E^{-1} T' + \Sigma_f^{-1} \mu_{F_{\text{prior}}}\right)$$

where $E = \Sigma$ is the deflection covariance (Hoey et al., 2013).

3. Action Selection and Utility Functions

BayesAct agents choose actions (propositional $A_t$ and affective $B^a_t$) to optimize a joint criterion at each step:

$$\arg\max_{a} \;\; \mathbb{E}\left[ -\| F - T \|^2 \right] + \beta\,\mathbb{E}[U]$$

or equivalently,

$$\arg\min_a\, \mathbb{E}[\|F-T\|^2] - \alpha\,\mathbb{E}[U]$$

with $\alpha,\beta$ determining the trade-off between affective coherence and instrumental reward $U$ (Hoey et al., 2019). The instantaneous reward is generally modeled as

$$R(S_t, B_t) = R_x(A_t, X_t) - \lambda\,\delta(F_{t+1}, T_{t+1})$$

where $R_x$ is task-oriented and $\lambda$ weights deflection cost (Hoey et al., 2013). Planning is performed using value iteration, Monte Carlo roll-out, or POMCP-style forward search, with belief updates interleaved (Hoey et al., 2013, Jung et al., 2017). In continuous EPA-action spaces, candidate affective actions $B^a_t$ are sampled from the normative (deflection-minimizing) distribution.

4. Dual-Process and Uncertainty-Weighted Reasoning

BayesAct supports a dual-process framework, dynamically balancing fast, low-dimensional “connotative” (affective) reasoning against full, context-sensitive “denotative” (task/logic) reasoning. The influence of each is weighted by the entropy of the denotative belief:

$$w_t = \frac{1}{1 + \exp\left[\kappa(H[s_t] - \theta)\right]}$$

Affective deflection then dominates in high-uncertainty settings, while task utility dominates in lower-uncertainty situations (Hoey et al., 2019).

5. Empirical Validation and Applications

BayesAct has been deployed and evaluated in multiple domains:

Networked Social Dilemmas

BayesAct agents in the Iterated Networked Prisoner’s Dilemma reproduce four of five hallmark properties of human play, including network invariance, anti-correlation of cooperation and reward, moody conditional cooperation, and player-type stratification. Classical imitation models do not replicate these human-like signatures (Jung et al., 2017).

Human-Computer Interaction

BayesAct-based tutoring systems and assistive agents adapt linguistic and behavioral affect to maintain low deflection, improving user satisfaction and effectiveness in tasks such as mathematics tutoring and dementia-care handwashing assistance. Empirical results show lower deflection and increased user-reported naturalness compared to non-Bayesian affective agents (Hoey et al., 2013).

Algorithmic Sketch

A canonical BayesAct infer-act loop proceeds as:

Initialize beliefs over F, T, X For t in 1...T: Observe denotative and emotional cues Belief update via particle filtering Compute expected deflection and utility Combine with uncertainty-weighted trade-off Plan/select action to maximize joint criterion Project action into affective/EPA space (if needed) Execute behavior and repeat

(Hoey et al., 2019, Hoey et al., 2013)

6. Extensions: Social-Psychological and Computational Phenomena

BayesAct’s Bayesian machinery supports flexible modeling of fairness, cognitive dissonance, and conformity through the dynamics of belief uncertainly and somatic (affective) coupling (Hoey et al., 2019):

• Fairness bias: High uncertainty in self-identity increases affective coherence pressure, reproducing empirically observed fairness effects.
• Cognitive dissonance and conformity: Bayesian updates of beliefs under affective feedback reproduce classic findings wherein perceived impressions and group influences modulate memory and judgment.

The model naturally generalizes well-known reinforcement learning exploration strategies: optimism under uncertainty (value bonus), random exploration, and affects both curiosity- and norm-driven behaviors as instances of deflection minimization (Hoey et al., 2019).

7. Limitations and Future Directions

Despite its principled integration of affective and instrumental reasoning, BayesAct is limited by the quality and resolution of underlying EPA lexica, the complexity of fully general POMDP inference in high-dimensional or multi-agent settings, and the challenge of mapping between low-dimensional affective representations and naturalistic social signals. Research directions include more expressive, learned EPA representations, robust affective observation models, and scalable planning for large-scale social systems (Hoey et al., 2013, Hoey et al., 2019).

In contrast, systems that use only deterministic ACT machinery for action conditioning—such as neural dialogue models leveraging ACT EPA outputs as fixed conditioning—do not constitute full BayesAct models, as they lack explicit Bayesian updating over beliefs, probabilistic observation models, and joint optimization of affect and utility (Asghar et al., 2020). The full BayesAct paradigm requires maintaining posterior distributions over latent affective states and integrating these with task-oriented planning.

---

References:

• "Affect Control Processes: Intelligent Affective Interaction using a Partially Observable Markov Decision Process" (Hoey et al., 2013)
• "‘Conservatives Overfit, Liberals Underfit’: The Social-Psychological Control of Affect and Uncertainty" (Hoey et al., 2019)
• "Socio-Affective Agents as Models of Human Behaviour in the Networked Prisoner's Dilemma" (Jung et al., 2017)
• "Generating Emotionally Aligned Responses in Dialogues using Affect Control Theory" (Asghar et al., 2020)