🤖 AI

A Computational Implementation of a Goal-Directed Theory of Affect

This paper presents the first high-fidelity computational implementation of the Goal-Directed Theory of affect, demonstrating how complex emotional dynamics naturally emerge from the interplay between discrepancy detection and action selection within a transparent, psychologically grounded agent framework.

Original authors: Bernhard Hilpert, Tamás Szűcs, Joost Broekens, Agnes Moors

Published 2026-09-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Bernhard Hilpert, Tamás Szűcs, Joost Broekens, Agnes Moors

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Human beings are constantly navigating a world of goals, from the simple desire to catch a bus to the complex ambition of building a career. For decades, scientists have tried to understand the invisible engine that drives these pursuits: emotion. In the field of affective computing, researchers build computer programs that can recognize or simulate feelings, hoping to create machines that interact more naturally with people. However, a long-standing tension has plagued this work. Many existing models treat emotions as a final label slapped onto a situation after the fact, like a stamp on a letter. Others rely on raw data signals that lack a deep connection to human psychology. The result is often a "black box" where a computer outputs a feeling without a clear, step-by-step explanation of how that feeling actually arose from the machine's internal processes.

A new study presented at the 2026 International Conference on Affective Computing and Intelligent Interaction seeks to resolve this by building a computer model based on a specific psychological idea called the Goal-Directed Theory. This theory suggests that emotion is not a separate event that happens after we think, but a functional byproduct of the very act of trying to achieve something. It proposes that our feelings emerge continuously as we detect the gap between where we are and where we want to be, and as we weigh our chances of closing that gap. The researchers, led by Bernhard Hilpert and his team, have translated this theory into a high-fidelity computer simulation. Their work demonstrates that complex emotional patterns, such as the rising hope of anticipation or the sudden drop of failure, can emerge naturally from simple mechanical interactions, without needing special, dedicated modules to generate them.

To test this idea, the team created a digital agent and placed it in two simple, controlled environments. The first was a virtual dice game. In this scenario, the agent had a clear goal: to roll a specific number, a six. The researchers programmed the agent to calculate its feelings based on two main factors. First, the agent measured the distance between its current state and its goal. If it was far away, it registered a negative feeling proportional to that distance. Second, the agent considered its expectations. If the agent knew it had a chance to roll the die, it calculated the probability of success. The simulation showed that even before the die was thrown, the mere presence of a viable action created a subtle "lift" in the agent's emotional state. This lift occurred because the agent had a path forward, even if the outcome was still uncertain. When the agent finally rolled the die, the feeling changed instantly. If it hit the target, the gap vanished, and the feeling spiked into a sharp positive peak. If it missed, the feeling dropped back to a negative baseline.

The researchers then introduced a second, more complex challenge: a virtual corridor. Here, the agent had to walk step-by-step toward the end of a hallway. Along the way, there was a small chance of falling into a trap at any step. This setup allowed the team to see how feelings evolved over time as the agent moved closer to its goal. They ran the simulation with different types of agents. Some were "oblivious," meaning they did not realize there was a trap and assumed they would reach the end with certainty. Others were "accurate," meaning they knew the exact probability of falling into a trap at each step. The results revealed a fascinating interplay between progress and expectation. As the accurate agent moved forward, its feeling of hope grew not just because it was physically closer to the finish line, but because the cumulative risk of falling into a trap decreased with every safe step. This created a non-linear rise in positive feeling, a "lift" that became stronger the closer the agent got to success.

In contrast, the oblivious agent, which did not perceive the changing risk, showed a flat emotional line until the very end. This highlighted a key finding of the study: the specific shape of an emotional experience depends entirely on how the agent processes its environment and its own chances of success. The simulations also captured the mechanics of failure. When the accurate agent finally fell into a trap, it experienced a "crash." This was not just a return to a negative state; it was a sudden, sharp drop caused by two things happening at once. The agent's progress toward the goal vanished, and its expectation of future success dropped to zero instantly. In the simpler, binary versions of the tasks where the agent either succeeded or failed without a middle ground, the feeling of failure was less dramatic but still distinct, driven by the total loss of any remaining hope.

The significance of this work lies in its transparency. Unlike previous models that might use complex, hidden rules to decide when an agent feels happy or sad, this new framework is built entirely on the visible mechanics of goal pursuit. Every emotional output in the simulation is a direct calculation of the gap between reality and a desired outcome, combined with the probability of bridging that gap. The researchers found that they did not need to program a separate "hope" or "fear" module. Instead, these complex emotional profiles emerged automatically from the basic math of discrepancy and expectation. This suggests that the feeling of anticipation is not a mysterious add-on to human behavior, but a natural consequence of having a goal and a plan to reach it.

While the study is currently limited to computer simulations and does not yet involve human participants, it establishes a powerful new tool for the field. The authors describe their model as a "computational laboratory" where different psychological theories can be tested and refined. By ensuring that every part of the computer code maps directly to a concept in human psychology, the researchers have created a system that is fully inspectable. This allows scientists to predict how an agent should feel in a specific situation before ever testing it on a real person. The work moves the field away from treating emotions as black-box heuristics and toward a granular, mechanical understanding where feelings are seen as integral parts of the process of acting in the world. It offers a clear path forward for building machines that do not just mimic human emotion, but generate it through the same fundamental processes that drive human behavior.

Technical Summary: A Computational Implementation of a Goal-Directed Theory of Affect

Problem Statement
Computational modeling of emotion has historically faced a tension between two approaches: descriptive, "snapshot-based" appraisal models that lack mechanistic granularity, and granular, signal-driven architectures (e.g., Reinforcement Learning) that often lack appropriate psychological grounding. Existing models often treat appraisal as a discrete event or a specialized layer mapping cognitive states to affective labels, failing to capture affect as a continuous, emergent property of an agent's internal processing cycles. Furthermore, many models rely on "black-box" heuristics rather than transparent, testable mechanisms derived directly from psychological theory.

Methodology
This paper presents the first high-fidelity computational implementation of the Goal-Directed Theory (GDT) of affect. The authors translate the psychological theory into a simulatable computational architecture where affect is not a post-hoc label but a functional byproduct of the interplay between discrepancy detection and action selection.

Theoretical Framework: The model defines an agent operating within a set of goal states ( $G$ ) and environmental states ( $S$ ). Affect is calculated through two distinct phases within a goal-directed cycle:
1. Discrepancy Detection ( $A^D$ ): Calculates the distance between a current state feature and a target goal value. Negative affect arises from the magnitude of the discrepancy weighted by goal importance; positive affect arises when the discrepancy is zero.
2. Action Selection ( $A^R$ ): Calculates the expected utility of available actions based on the probability of reducing the discrepancy. This phase generates an affective trace based on the presence and utility of viable behavioral options.
3. Integration: Total affect ( $A^{total}$ ) is the sum of the discrepancy-derived affect and the action-selection-derived affect.
Experimental Design: The model was evaluated using two controlled simulation tasks designed to isolate specific affective signatures:
1. Dice Task: A single-step task to isolate the affective signatures of discrepancy and expectancy. It included "Binary" (hit-or-miss) and "Gradual" (proximity-based) versions.
2. Corridor Task: A multi-step sequential task introducing temporal dynamics. It combined Binary/Gradual discrepancy versions with "Oblivious" (constant expectancy) and "Accurate" (discounted probability of traps) agent types.

Key Contributions

High-Fidelity Implementation: The paper provides the first computational realization of the GDT, mapping every computational component directly to a psychological construct (e.g., goal discrepancy, expected utility) without introducing external heuristics.
Emergent Affective Profiles: The study demonstrates that complex affective dynamics, such as an anticipatory "lift" and a failure "crash," emerge naturally from the interaction of simple discrepancy and utility calculations. No dedicated "anticipation" or "emotion" modules were required.
Mechanistic Transparency: The framework serves as a "computational laboratory," allowing researchers to test specific functional forms of psychological hypotheses. It moves the field away from opaque, top-down rules toward a granular, mechanistic understanding of how affect integrates with core agent behavior.
Simulation-Empiry Loop: By formalizing the theory into a simulatable model, the authors establish a framework for a continuous research loop where computational predictions can guide and be validated by future empirical studies with human participants.

Results
The simulations successfully reproduced distinct affective profiles predicted by the GDT:

Anticipatory Lift: In the Corridor task, agents with accurate expectancies showed a non-linear increase in total affect ("lift") as they approached the goal, even before the discrepancy was fully resolved. This was driven by the increasing expected utility of success ( $A^R$ ).
Gradual vs. Binary Dynamics: In "Gradual" versions, total affect scaled linearly with proximity to the goal. In "Binary" versions, affect remained stagnant until the final step, highlighting the role of granularity in discrepancy detection.
Failure "Crash": Upon hitting a trap or failing a goal, the model exhibited a "crash" characterized by a simultaneous drop in both components: $A^D$ plummeted to a failure baseline (in gradual versions), and $A^R$ dropped to zero due to the loss of viable action options.
Component Isolation: The Dice task confirmed that $A^D$ is driven purely by state-goal distance, while $A^R$ is driven by the existence of viable actions, and their sum creates the total affective state.

Significance and Claims
The authors claim that this work bridges the gap between high-level cognitive appraisals and low-level active processing. By ensuring that affect arises directly from the agent's internal dynamics (discrepancy and action selection), the model offers a parsimonious account that avoids the functional distinction often found in other architectures between "core cognition" and "affect."

The paper modestly asserts that this architecture provides a transparent, testable framework rather than a final, complete theory of emotion. It emphasizes that the current implementation relies on simplifying assumptions (e.g., static transition probabilities, simple subtraction for discrepancy, no temporal decay) and that the primary value lies in establishing a "computational toolbox." This tool allows for the systematic disentanglement of affective components and facilitates a "simulation-empiry" loop, enabling researchers to predict how experimental manipulations will manifest in affective profiles before conducting human studies. The work does not claim to replace existing models but to provide a mechanistic foundation for moving beyond "black-box" heuristics toward a granular understanding of affect.

Technical Summary: A Computational Implementation of a Goal-Directed Theory of Affect

More like this