Computational Models of Emotion¶
Overview of the main computational models of emotion, with a critical analysis of their strengths, limitations, and relevance to our "Affective Field Theory" approach.
1. Dimensional Models¶
Russell — Circumplex Model of Affect (1980)¶
James Russell proposed that all affective states can be mapped onto a two-dimensional circular space: - X axis: Valence (pleasure — displeasure) - Y axis: Arousal (activation — deactivation)
Labels for discrete emotions (happy, sad, angry, calm, etc.) are distributed around this circle as regions in continuous space, not as discrete points.
HIGH AROUSAL
|
excited | alarmed
elated -------- +A -------- afraid
happy | annoyed
--------- +V ---+--- -V ---------
content | miserable
serene -------- -A -------- sad
calm | depressed
LOW AROUSAL
Strengths: - Continuous: no artificial discretization - Well validated empirically across cultures - Computationally efficient (2 coordinates) - Captures the fundamental structure of affect
Limitations: - Lacks the Dominance dimension (anger and fear have similar valence/arousal but opposite dominance) - Does not capture dynamics (it is a state model, not a process model) - Some complex emotions are difficult to position
Relevance to the library: The circumplex provides the substrate for Layer 1 — Core Affect. Every affective state of the system has coordinates (valence, arousal) that position it in the circumplex.
Mehrabian & Russell — PAD Model (1974)¶
Extends the circumplex to three dimensions: - Pleasure (pleasant — unpleasant): ≈ valence - Arousal (excited — calm): ≈ arousal - Dominance (dominant/in-control — submissive/controlled)
The third dimension Dominance distinguishes emotions that are conflated in the 2D circumplex: - Anger: high arousal, low valence, high dominance - Fear: high arousal, low valence, low dominance
The PAD model claims these three dimensions are necessary and sufficient to characterize any emotional state.
Empirical validation: Mapping hundreds of emotion terms from 50+ languages onto these three axes shows significant cross-cultural consistency.
Relevance: The PAD model is the natural candidate for Core Affect representation in our architecture — three continuous axes covering the affective space with sufficient granularity without over-engineering.
2. Evolutionary Discrete Models¶
Plutchik — Wheel of Emotions (1980)¶
Robert Plutchik proposed a hybrid evolutionary-dimensional model.
Structure: - 8 primary emotions in opposing pairs: Joy/Sadness, Trust/Disgust, Fear/Anger, Surprise/Anticipation - Arranged in a 3D cone with intensity on the vertical axis (e.g. annoyance → anger → rage) - Dyads (blending): adjacent emotions combine (e.g. joy + trust = love; fear + surprise = awe)
Ecstasy
Joy Admiration
Serenity Trust
Anticipation Acceptance
Vigilance Apprehension
Anger Fear
Disgust Surprise
Loathing Amazement
Grief
Sadness
Pensiveness
Evolutionary function of each emotion: - Fear → Protection - Anger → Destruction of barriers - Joy → Reproduction - Sadness → Reintegration - Trust → Affiliation - Disgust → Rejection - Surprise → Orientation - Anticipation → Exploration
Strengths: - Supports emotional blending - Intensity dimension - Evolutionary grounding - Intuitive and visually communicable
Limitations: - The choice of 8 primitives is arbitrary - Blending is additive; it does not capture the complexity of emotional emergence
Relevance: The Wheel provides a natural taxonomy for the discrete emotion categories that emerge from the continuous PAD space. The regions of Plutchik's cone can be mapped onto PAD space, providing human-readable labels for numerical coordinates.
Panksepp — 7 Primary Affective Systems (1998)¶
The seven emotional command systems identified by Panksepp through experimental neurobiology
(see 02_neuroscience.md for details):
| System | Approx. PAD | Behavior |
|---|---|---|
| SEEKING | P+, A+, D+ | Exploration, curiosity |
| RAGE | P-, A+, D+ | Aggression, frustration |
| FEAR | P-, A+, D- | Avoidance, freeze/flight |
| LUST | P+, A+, D+ | Sexual behavior |
| CARE | P+, A-, D+ | Nurturance, attachment |
| PANIC/GRIEF | P-, A-, D- | Isolation, crying |
| PLAY | P+, A+, D- | Social bonding, play |
Relevance: These 7 systems can function as primary clusters in PAD space — regions around which basic emotional responses organize. Biologically more robust than Plutchik's Wheel, but less granular for linguistic use.
3. Cognitive and Appraisal Models¶
OCC Model — Ortony, Clore & Collins (1988)¶
The most influential computational model of emotion in symbolic AI. Generates 22 emotion types from appraisals of three categories:
- Events → pleased/displeased (e.g. joy, distress, hope, fear, satisfaction, disappointment)
- Agents (actions) → approving/disapproving (e.g. pride, shame, admiration, reproach)
- Objects → liking/disliking (e.g. love, hate)
Compound emotions emerge from combinations (e.g. gratification = joy + pride; remorse = distress + shame).
EMOTIONS
/ | \
Events Agents Objects
/ \ / \ / \
pleased disp. app. repr. like dislike
| |
hope distress joy fear ...
Strengths: - Rule-based → interpretable and transparent - Compositional structure - De facto standard for virtual agents (NPCs in video games, social chatbots)
Limitations: - Discrete → loses the continuity of emotional experience - Rule-based → rigid, does not learn - Does not model temporal dynamics - Difficult to scale to open-ended scenarios
Relevance: OCC provides a structured vocabulary of emotion labels that can be used as a semantic layer on top of PAD coordinates. An event with (valence=+0.7, arousal=0.4) in response to a goal-furthering event → OCC label "joy" or "satisfaction".
Scherer — Component Process Model (CPM)¶
(Detailed in 02_neuroscience.md — section 5)
The CPM is the richest among appraisal models. The sequence of 5 SECs (Stimulus Evaluation Checks) generates emotional states in a processual way.
Computational format (original Scherer CPM):
appraisal_result = {
novelty_check: float, # 0=completely expected, 1=entirely novel
intrinsic_pleasantness: float, # -1=unpleasant, +1=pleasant
goal_significance: float, # -1=obstructs, +1=furthers
coping_potential: float, # 0=no coping, 1=full control
norm_compatibility: float # -1=violates norms, +1=conforms
}
These 5 values determine the resulting emotional state through compositional rules.
Relevance: The CPM is the model that most directly informs Layer 4 — Appraisal Vector in our architecture.
Implementation note: In the library, intrinsic_pleasantness is replaced by self_relevance
(how much the event concerns the system's "self", [0.0, 1.0]). This deviation from the original
CPM is intentional: self_relevance is more operationalizable in an LLM agent and contributes to
arousal (not to valence), preserving the 5 SECs while redistributing their roles.
4. Architectural Cognitive Models¶
ACT-R — Adaptive Control of Thought—Rational (Anderson, 1983+)¶
ACT-R is a unified cognitive architecture. Retrieval is governed by activation levels:
Where:
- Bi = base-level activation (recency and frequency with power-law decay)
- Σ(Wj * Sji) = spreading context activation
- εi = Gaussian noise
Base-level decay:
Wheret_j are the elapsed times since each access and d is the decay parameter.
Emotional extensions of ACT-R: Researchers have proposed extensions where:
- Emotion modifies the d parameter (decay) — emotionally intense memories decay more slowly
- Emotion modifies Wj — the current emotional context gives greater weight to emotionally
congruent memories
Relevance: ACT-R's base-level activation mechanism is the foundation for our differentiated decay system. We adopt the power law but add modulation by arousal.
SOAR with Emotion (Marinier, Laird & Lewis, 2009)¶
"A Computational Unification of Cognitive Behavior and Emotion" integrates appraisal theory into SOAR. Emotional evaluations serve as intrinsic rewards for reinforcement learning.
Phases: - Relevance appraisal during Perceive/Encode - Evaluative appraisal during Comprehend - Outcome probability during Intend
Relevance: The emotion-as-intrinsic-reward principle is applicable to our architecture: emotions not only tag memories but also modulate the reinforcement of system behavior.
CLARION — Motivational Subsystem (Sun, 2016)¶
CLARION includes an explicit motivational subsystem with: - Low-level drives (survival, curiosity) - High-level drives (purpose, focus, adaptation)
Emotion is generated by drive activations and action potentials, with appraisal and metacognition as secondary factors.
Relevance: The Spinozian conatus in our architecture corresponds to CLARION's drives — the system's goals that generate the evaluative plane of appraisal.
5. Recent Neural-Symbolic Models (2023–2026)¶
Chain-of-Emotion Architecture (2024)¶
Architecture for affective LLM agents in video games. Uses a separate LLM call for emotional appraisal before every response:
Evaluated as significantly more natural and emotionally sensitive than controls (PLOS One, 2024).Relevance: The two-pass structure (first appraisal, then response) is the pattern we adopt in our appraisal engine.
Emotional RAG (Huang et al., 2024)¶
Two mood-dependent retrieval strategies: 1. Combination: joint weighting of semantics + emotional state 2. Sequential: filtering by emotion then semantic ranking (or vice versa)
Outperforms standard RAG on role-playing datasets.
Relevance: Confirms the utility of the emotional_congruence signal in retrieval scoring.
Our approach extends this with momentum, the Mood field, and spreading activation.
Dynamic Affective Memory Management (2025)¶
Bayesian memory updating with entropy minimization. Uses memory entropy as a quality metric — the system minimizes global entropy through Bayesian updates to the affective memory vector.
Relevance: The entropy-as-memory-quality principle is promising for our consolidation
system. More precise memories (low entropy) should have higher consolidation_strength.
6. Comparative Analysis: Why No Existing Model Is Sufficient¶
| Model | Static/Dynamic | Continuous/Discrete | Appraisal | Temporal | Global |
|---|---|---|---|---|---|
| Russell Circumplex | Static | Continuous | No | No | No |
| PAD | Static | Continuous | No | No | No |
| OCC | Process | Discrete | Yes | No | No |
| Plutchik Wheel | Static | Hybrid | No | Yes (intensity) | No |
| Panksepp 7 | Static | Discrete | No | No | No |
| Barrett Core Affect | Static | Continuous | Partial | No | No |
| CPM (Scherer) | Process | Continuous | Yes | Partial | No |
| ACT-R ext. | Dynamic | Continuous | No | Yes | No |
| Emotional RAG | Process | Hybrid | No | No | No |
Systematic gaps identified: 1. No model captures the temporal dynamics of affect (where you are going, not just where you are) 2. No model has a global field (mood field) that colors all operations 3. No model integrates appraisal + memory + retrieval in a unified framework 4. No model implements reconsolidation as a native operation
These four gaps motivate our Affective Field Theory.
Bibliographic Notes¶
- Russell, J.A. (1980). "A circumplex model of affect." Journal of Personality and Social Psychology, 39, 1161-1178.
- Mehrabian, A. & Russell, J.A. (1974). An Approach to Environmental Psychology. MIT Press.
- Plutchik, R. (1980). Emotion: A Psychoevolutionary Synthesis. Harper & Row.
- Panksepp, J. (1998). Affective Neuroscience. Oxford University Press.
- Ortony, A., Clore, G.L. & Collins, A. (1988). The Cognitive Structure of Emotions. Cambridge University Press.
- Anderson, J.R. et al. (2004). "An integrated theory of the mind." Psychological Review, 111(4), 1036-1060.
- Marinier, R.P., Laird, J.E. & Lewis, R.L. (2009). "A computational unification of cognitive behavior and emotion." Cognitive Systems Research.
- Sun, R. (2016). Anatomy of the Mind. Oxford University Press.
- Huang et al. (2024). "Emotional RAG: Enhancing Role-Playing Agents through Emotional Retrieval." arXiv:2410.23041.
- Zhang & Zhong (2025). "Decoding Emotion in the Deep." arXiv:2510.04064.
- Dynamic Affective Memory Management (2025). arXiv:2510.27418.