Learning Emotional Dynamics: A World Model Approach

Traditional emotion recognition systems perform frame-wise classification of af-fective states, ignoring the rich temporal structure inherent in emotional experi-ences. This paper presents a novel framework for modeling emotional states asdynamical systems using world models adapted from reinforcement learning. Ourapproach learns compressed representations of physiological signals (blood volumepulse, electrodermal activity, and skin temperature) and predicts their temporalevolution through transformer-based dynamics modeling. We systematically com-pare three conditioning strategies—physiological signals, emotion labels, and ran-dom baseline—to establish causal relationships between information sources andprediction accuracy. Experiments on the WESAD dataset (N = 15 subjects, 3emotional states) demonstrate that physiological conditioning achieves 52.5% er-ror reduction compared to random baseline and outperforms label conditioning by34.1%. Multi-step rollout analysis reveals accurate predictions up to 5 timestepsahead (R2 = 0.74), suggesting a characteristic temporal horizon for emotional dy-namics. Optimal latent dimensionality (k = 5) balances classification accuracy(76% F1-score) with reconstruction fidelity (MSE=6.7 × 10−3). Our findings chal-lenge the dominant frame-wise paradigm in affective computing by demonstratingthat emotions exhibit learnable, structured dynamics, enabling not only recognitionbut forecasting of emotional states.Keywords: Affective Computing, World Models, Physiological Signals, Emotion Fore-casting, Causal Analysis, Deep Learning, Transformers, Time Series Prediction