Affective States as Motivational Engines for Action Selection

45 concepts35 papers2 proposals

In traditional artificial intelligence, agents are often designed as "symbol pushers" that execute logic to achieve externally defined goals. However, as William S. Faught argued as early as 1975, true autonomy requires an internal motivational engine. This engine is affect: a system of internal states that assigns subjective significance to thoughts and actions based on an agent's self-interest.

Rather than treating emotion as a decorative layer, researchers in this field view affective states as functional control mechanisms. These states bridge the gap between raw physiological needs and high-level cognitive planning, ensuring that an agent's behavior remains coherent, adaptive, and intrinsically driven.

From Task-Following to Value-Driven Autonomy

Most autonomous agents react to external task lists. In contrast, a "value-driven" architecture prioritizes internal needs. Chen et al. demonstrate that agents imbued with intrinsic values and physiological needs (like hunger or social connection) exhibit more believable and consistent long-term behavior than those following predefined rules.

This shift moves the agent from a reactive state to a proactive one. As Wang et al. show, agents driven by internal "desires" can autonomously generate natural sequences of daily activities—such as choosing to rest when fatigued rather than simply waiting for the next command. This mirrors animal ethology; Maes et al. found that incorporating internal needs and sensory fatigue prevents the "dithering" common in purely logical systems, where an agent might rapidly switch between two equally logical actions without ever completing one.

Affect as a Computational Heuristic

In a complex environment, an agent faces a combinatorial explosion of possible actions. Affective states serve as a heuristic to prune this search space. Usef Faghihi proposes assigning "emotional valences" to information units, allowing the agent to prioritize memories and experiences that have high perceived significance or past payoff.

This prioritization often takes the form of a weighted voting process. Margaret Manella Kozak describes an architecture where internal feelings cast votes for candidate actions. The intensity of the feeling determines the weight of the vote, leading to non-deterministic behavior. For example, a high "fear" state might heavily weight withdrawal actions, overriding a "curiosity" drive that would otherwise lead the agent toward a novel stimulus. Similarly, Yu et al. utilize a multi-stage framework where predicted actions are filtered through the agent's current emotional state, ensuring that individual choices align with a broader psychological context.

The Emergence of Functional Emotion

A key insight in this field is that emotions are not "add-ons" but emergent properties of intelligent systems managing conflicting motives. Sloman and Croucher argue that any agent with multiple goals in an unpredictable world will inevitably exhibit "emotions" as its internal administrative processes interrupt and re-prioritize tasks.

This functional view allows complex states like hope and fear to be modeled mathematically. Moerland et al. demonstrate that these states emerge naturally from model-based reinforcement learning:

  • Hope emerges from the anticipation of potential future rewards during planning.
  • Fear emerges from the evaluation of potential risks or negative outcomes.

By deriving these states directly from the agent's planning machinery, they become integral to the decision-making process rather than arbitrary variables.

Durative States and Behavioral Momentum

Human emotions are "durative"—they linger and color future reactions rather than vanishing the moment a stimulus is gone. Tanguy et al. emphasize that for AI to be effective, emotional states must have momentum. A lingering state of "anger" or "joy" should influence the stylistic execution of actions and the selection of subsequent goals long after the triggering event.

This persistence provides a form of architectural stability. As Li et al. suggest, modeling the dynamic competition between internal motivation and emotional state allows for adaptive decision-making. When an agent's internal teaching drive influences its emotional state, it can adjust its tutoring strategy based on both its own performance and the perceived engagement of the student, creating a feedback loop that mimics the nuanced social-emotional regulation found in the human prefrontal cortex and amygdala (Xue Li).

Ultimately, this research direction suggests that "intelligence" cannot be separated from "affect." By treating emotions as internal control signals that weight rewards, filter memories, and manage motives, we move closer to agents that possess true functional agency.

Key Foundational Work

Affect As Motivation For Cognitive And Constive (1975) — contributed 4 concept elements to this direction.
Affect As Motivation For Cognitive And Conative Processes (1975) — contributed 1 concept elements to this direction.
TALE-SPIN, An Interactive Program that Writes Stories (1977) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_affective_action_selection_3ea5beda,
  title = {affective action selection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/affective-action-selection/rdr_3ea5bedacd9c434e9a8fb3472a775807},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_affective_action_selection_3ea5beda_v1,
  title = {affective action selection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/affective-action-selection/rdr_3ea5bedacd9c434e9a8fb3472a775807/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake