A Computational Implementation of a Goal-Directed Theory of Affect

Summary

The paper presents what the authors call the first computational implementation of the Goal-Directed Theory (GDT) of affect, a response-evaluation theory in which affect arises from goal-discrepancy detection and action selection rather than appraisal snapshots. The formalization (Eqs. 1–17) defines goal states, feature-based discrepancies, action expectancies, and expected utilities, and computes two affect traces—discrepancy-detection affect (A^D) and action-selection affect (A^R)—whose sum yields total affect. Using parsimonious default functions where the theory is silent, the authors simulate a single-step Dice task (binary vs. gradual goals) and a multi-step Corridor task (binary/gradual discrepancies × oblivious/accurate expectancies). The four Corridor conditions yield distinct profiles: flat baselines, linear rises from shrinking discrepancy, non-linear 'expectancy lifts' from compounded transition probabilities (0.9^n), and a pre-goal peak; failure produces a 'crash' separating loss (A^D drop) from disappointment (A^R drop). Anticipation thus emerges without dedicated modules. Limitations are acknowledged: toy tasks only, no learning dynamics, no temporal decay, single-goal scope, unresolved scale mismatch between A^D (spanning 0–5) and A^R (spanning 0–1), and no comparison against alternative affect models or human data; empirical validation is deferred to future work.