PaperlayerView exact revision on arXiv

Source-linked AI summary

A Computational Implementation of a Goal-Directed Theory of Affect

Bernhard Hilpert, Tamás Szűcs, Joost Broekens, Agnes Moors

arXiv:2609.06654v1cs.AI

TL;DR

The paper addresses the need for mechanistically granular computational models of affect by implementing the Goal-Directed Theory in a simulatable framework. Across Dice and Corridor simulations, affective profiles such as anticipatory lift and failure crash emerge from interactions between goal discrepancy and action-selection expectancies.

  • Problem

    Computational implementations of affect require mechanistic mappings between appraisal processes and affective outcomes that existing approaches do not always specify in sufficient detail.

  • Method

    The paper formalizes the GDT as a generalized, scalable computational architecture and evaluates it through Dice and Corridor simulations.

  • Results

    Affective profiles, including anticipatory lift and failure crash, emerged as functional byproducts of the joint dynamics of goal discrepancy and expected utility without dedicated affect modules.

  • Takeaways & Limitations

    The model provides a transparent computational toolbox for inspecting GDT mechanisms and supporting a simulation-empiry loop before empirical studies.

  • Takeaways & Limitations

    The evaluation uses intentionally simple Dice and Corridor tasks to facilitate comparison with human experimental paradigms and empirical validation.

Abstract

from arXiv · show

Computational modeling of emotion has long faced a tension between descriptive, "snapshot-based" appraisal models and granular, signal-driven architectures that often lack appropriate psychological grounding. This paper addresses this gap by presenting the first high-fidelity computational implementation of the Goal-Directed Theory (GDT) of affect. In this framework, affect is not a post-hoc label but a functional byproduct emerging from the continuous interplay between discrepancy detection and action selection within an agent's internal processing cycles. We evaluate the model through a series of principled simulations (Dice/Corridor tasks) designed to isolate affective signatures and dynamics during multi-step goal pursuit. Results demonstrate that complex affective profiles, like an anticipatory "lift" and a failure "crash", emerge naturally from simple interactions between goal-discrepancy and action-selection expectancies without requiring additional dedicated modules. By ensuring every computational component maps directly to components of the psychological theory, this work establishes a transparent, testable framework that enables a continuous "simulation-empiry" research loop. Our work contributes to moving the field beyond "black-box" heuristics toward a granular, mechanistic understanding of affect, integrated into the core of agent behavior.

I. INTRODUCTION

The paper motivates a computational implementation of the Goal-Directed Theory (GDT) because it specifies affect generation with greater mechanistic granularity than many appraisal accounts. It introduces a model in which affect emerges during goal-directed cycles from discrepancy detection, action selection, and expected outcomes.

  • Motivation: Appraisal theories do not always specify mappings between appraisal patterns and affective outcomes with the mechanistic granularity required for computation.The GDT instead describes affect generation through underlying principles and processing cycles.
  • Contribution: The paper presents the first computational implementation of the GDT and aims to support future comparison between simulated affective profiles and human affect patterns.Its stated goals are high-fidelity translation of the theory into a simulatable model and an experimental loop between modeling and empirical validation.
  • Theoretical framework: Unlike stimulus-evaluation appraisal theories, the GDT predicts affect and other emotion components from stimulus–goal discrepancies and action-selection processes.The GDT is characterized as a response-evaluation theory, whereas appraisal theories predict affect from appraisal patterns.
  • Theoretical framework: When a discrepancy exists, the cycle activates a second goal to reduce or eliminate it through accommodation, immunization, or assimilation.If assimilation is selected, action options are compared and the option with the highest expected utility is chosen.
  • Theoretical framework: The chosen action generates a predicted outcome, and the resulting stimulus is fed into the next cycle for comparison with the valued and expected outcome.Any discrepancy in this comparison can function as a reward prediction error for updating the chosen action option's expectancy.
  • Theoretical framework: Affect traces arise from discrepancy detection and action selection across parallel goal-directed cycles and combine into the individual's total affective state.The figure describes discrepancy-based polarity and intensity alongside action-option utility as sources of affective traces.

B. Existing computational models

Existing computational models either map appraisals to affect descriptively or separate appraisal from core cognition, whereas the GDT makes affect an integral product of continuous goal-directed processing. The implementation formalizes this account through discrepancy detection, action-selection expectancies, and their integration into total affect.

  • Traditional architectures often use predefined rules to map situational appraisals to affective labels, leaving generation largely descriptive.
  • The GDT links affect directly to the same discrepancy-detection and action-selection processes that drive behavior.
  • The implementation uses parsimonious default functions while keeping functional forms open for future psychological hypothesis testing.The authors describe the framework as a computational laboratory for comparing alternative functions.
  • The model represents goals as target feature values with importance weights, states as feature-score vectors, and actions as state-specific responses.
  • Affect is computed through discrepancy and expected-utility traces, which are combined into total affect for each state.Discrepancies compare target values with state features; action expectancies determine expected utility.
  • The current implementation assumes state certainty, omits serial dependence and temporal decay, and runs a complete goal-directed cycle in each state.These are simplifying choices that delimit the present model’s temporal and representational scope.

A. Simulation Framework & Objectives

The evaluation uses controlled simulations to isolate how stimulus–goal discrepancy and action-selection expectancy contribute to affect. It compares a single-step Dice task with a sequential Corridor task across binary or gradual progress and oblivious or accurate expectancies.

  • The simulations focus on the interaction between stimulus–goal discrepancy and action-selection expectancies.
  • The Dice task isolates the two affective components in a single-step setting.
  • The Corridor task introduces sequential expectancy effects during multi-step goal pursuit.
  • Both tasks vary goal progress between binary and gradual versions and action expectancies between oblivious and accurate versions.

Task 1: Dice task

The Dice task models an agent seeking a six through probabilistic die throwing, contrasting binary hit-or-miss features with gradual proximity to the target. Its state diagram captures transitions from starting and throwing states through outcome states and back to the next trial.

  • From sthrow, the sole throwing action leads to each of six outcome states with probability 1/6.
  • The agent seeks the goal state representing a die roll of six, with goal importance v1 = 1.
  • At sstart, the agent has discrepancy but no goal-conducive action, producing negative affect from discrepancy detection alone.
  • The binary version makes only the six outcome discrepancy-free, whereas the gradual version represents proximity to the target through feature scores.
  • The state diagram distinguishes V1 binary and V2 gradual feature scores, with the rightmost transition returning to sthrow for the next trial.

Task 2: Corridor task

The Corridor task models sequential goal pursuit through probabilistic state transitions, contrasting binary versus gradual discrepancies and oblivious versus accurate expectancies. These combinations generate distinct affect profiles from the interaction of discrepancy detection and expectancies.

  • Task setup: The agent moves rightward through corridor states with a 0.9 transition probability and enters a trap with probability 0.1.The goal is the corridor’s rightmost state, while the trap represents failure.
  • Expectancy modeling: Sequential expectancy is modeled by discounting subjective transition probabilities across the steps required to reach the goal.The model introduces stepwise transition probabilities to represent the probability of eventually reaching the goal state.
  • Experimental conditions: Four Corridor versions combine binary or gradual discrepancies with oblivious or accurate subjective transition probabilities.Oblivious agents assume perfect transition probabilities, whereas accurate agents align subjective probabilities with the true 0.9 and 0.1 risks.
  • Predicted affect profiles: Binary discrepancies produce flat AD, gradual discrepancies produce linearly increasing AD, and accurate expectancies produce a non-linearly increasing AR component.Oblivious expectancies remain flat, yielding four distinct affect profiles from the two dimensions.

V. RESULTS

The Dice and Corridor simulations show that affective intensity changes according to perceived goal discrepancy and action expectancies during both single-step and multi-step pursuit. Goal achievement produces a peak, while failure can produce a sharp affective crash or an expectancy-driven drop.

  • Dice task: In the binary Dice task, affect remains at a negative baseline for non-goal outcomes and spikes sharply when the goal is achieved.The spike is driven by elimination of discrepancy, with Dij = 0 at the goal.
  • Dice task: In the gradual Dice task, higher rolls produce higher affective intensity because they reduce numerical distance to the target.Near-miss outcomes therefore yield higher intermediate affect than distant outcomes.
  • Corridor task: In the binary-oblivious Corridor condition, AD and AR remain horizontal, leaving total affect at a negative baseline until the goal or trap state.The agent perceives neither goal proximity nor changing trap risk.
  • Corridor task: Accurate expectancies create a subtle upward affective trend in the binary Corridor condition as perceived cumulative trap risk decreases.AD remains flat, while AR increases as the agent moves farther from the start.
  • Corridor task: The gradual-accurate Corridor condition produces the highest affective intensity before success through combined linear discrepancy reduction and non-linear expectancy lift.Across versions, reaching the goal yields a peak with Dij = 0 and AR = 1.
  • Failure dynamics: Failure in gradual conditions produces a crash in both components, whereas binary failure lowers total affect through AR falling to zero.In binary versions, AD stays at its negative baseline while the loss of viable success expectancy drives the drop.

VI. DISCUSSION

The simulations support the GDT claim that total affect is a predictable joint outcome of goal discrepancy and expected utility. They show that anticipatory lift and complex affective profiles emerge from these components within the same computational framework.

  • Core interpretation: Total affect is a predictable joint outcome of goal discrepancy and expected utility across the Dice and Corridor tasks.The model examines each component’s contribution and their dynamics in computo.
  • Anticipatory affect: Non-zero expectancies provide an anticipatory lift to total affect even when goal discrepancy remains negative.Viable behavior options create this lift in the Dice task, while the Corridor task extends it to more complex profiles.
  • Affective dynamics: The Corridor simulations show that combining gradual discrepancies with compounded expectancies produces linearly and non-linearly increasing affective components.In the current implementation, total affect is mainly driven by unscaled discrepancy, with a smaller contribution from expected utility.

A. Limitations and Implications

The framework’s simulations show distinct affective dynamics during goal pursuit, while its current formulation retains simplifying assumptions and requires empirical validation. Its scalable architecture is intended to support iterative refinement and future simulation–empiry research.

  • A. Limitations and Implications: Gradual discrepancy produces linearly increasing total affect as the agent approaches the target in the oblivious Corridor task.The figure includes trials ending in traps at steps 4 or 5 and a trial reaching the goal.
  • A. Limitations and Implications: The current framework uses simplifying assumptions about functional forms because the GDT makes few specific predictions about them.Future work is intended to test alternative functions against empirical data.
  • A. Limitations and Implications: The scalable architecture allows researchers to refine specific formulas without rewriting the underlying framework.This supports a simulation–empiry loop in which affective profiles can be examined before human experiments.
  • A. Limitations and Implications: Accurate expectancy combines with gradual discrepancy to produce peak affective intensity before the agent reaches the goal.The non-linear expectancy lift interacts with the linear reduction in discrepancy.

VII. CONCLUSION

The paper presents a high-fidelity computational implementation of the GDT that maps psychological mechanisms into a simulatable framework. It shows that complex affective profiles can emerge from goal discrepancy and action selection without dedicated affect modules.

  • VII. CONCLUSION: The model systematically separates the contributions of goal discrepancy and action selection to affective experience.Its components map to states that can be controlled and measured in human subjects.
  • VII. CONCLUSION: Anticipatory affective lift emerges naturally as a functional byproduct of active behavior without requiring a separate dedicated module.The result is presented as evidence that affect can arise from the mechanisms generating behavior.
  • VII. CONCLUSION: The framework functions as a transparent computational toolbox for evaluating and extending the GDT.Its granular architecture connects high-level cognitive appraisals with low-level active processing.
  • VII. CONCLUSION: The implementation supports a continuous simulation–empiry loop for comparing computational affect signatures with human measurements.The proposed loop is enabled by mapping formalized components to measurable human states.

ETHICAL IMPACT STATEMENT

The study frames its impact as foundational and scientific, emphasizing transparent simulation, reproducibility, and explainability. It also identifies risks of emotional manipulation and states boundaries on clinical use and human validation.

  • ETHICAL IMPACT STATEMENT: The computational laboratory enables controlled, reproducible prediction of human affective responses without involving human participants or sensitive personal data.The current study relies entirely on principled computational simulations and documents its materials publicly.
  • ETHICAL IMPACT STATEMENT: Psychologically grounded affective profiles could support socially intelligent agents, but such systems could also be used to manipulate user emotions.The paper advocates prioritizing user autonomy and emotional well-being.
  • ETHICAL IMPACT STATEMENT: Mechanistic transparency is intended to make the triggers of agent affect auditable and mitigate risks associated with opaque models.The architecture favors granular interpretability over black-box affect generation.
  • ETHICAL IMPACT STATEMENT: The model is a foundational theory rather than a clinical tool and should not support diagnosis or treatment without extensive validation and ethical oversight.The paper limits clinical use because the model’s relevance to mental-health concepts does not establish clinical validity.
  • ETHICAL IMPACT STATEMENT: Future validation against human empirical data is to be conducted under informed-consent and participant-privacy protections.The framework maps computational components to measurable human states, but human validation remains future work.
Loading 2609.06654v1…