Abstract
This paper introduces the emotivector, a lightweight anticipatory mechanism designed to enhance the believability of synthetic characters. By predicting future sensor states and measuring the discrepancy between expectation and reality, the system generates emotional responses and focus-of-attention. This approach moves anticipation from a secondary role in planning to a primary driver of behavior at the sensor level, allowing for emergent human-like reactions without complex top-down affective modeling.
Introduction
Creating "believable" AI characters—those that successfully maintain a user's suspension of disbelief—remains a significant challenge in game design and autonomous agent research. Traditional methods often oscillate between purely artistic animation-focused approaches and complex psychological modeling. This paper argues that anticipation is the missing link. While most AI is reactive, human behavior is deeply rooted in expectations; we experience surprise, relief, or disappointment based on whether reality aligns with our predictions.
The Emotivector Architecture
The proposed architecture centers on the emotivector, a modular component attached to an agent's sensors that provides a sense of "gut feeling" through anticipation.
Prediction and Learning
Each emotivector utilizes a hybrid algorithm inspired by the Kalman filter and the generalized recirculation algorithm. It maintains a history of sensor data to predict the next state ($x_t$) and employs a "delta rule" to update its internal model based on prediction errors.
Percept Salience and Sensation
The system calculates the "interest" or salience of a signal based on exogenous factors (surprise) and endogenous factors (goal-seeking). Furthermore, it maps the mismatch between expected and actual rewards or punishments to four primary sensations:
- S+ (Positive Increase): Better than expected.
- $+ (Positive Reduction): Good, but not as good as expected.
- S- (Negative Increase): Worse than expected.
- $- (Negative Reduction): Bad, but not as bad as expected.
To handle uncertainty, a meta-predictor tracks the error of the primary predictor. If the prediction error exceeds the meta-predictor's expectation, the signal is flagged as highly relevant, effectively modeling "trust" in the environment.
Experimental Evaluation
The model was tested using a custom 3D Word Puzzle Game. 280 puzzles were played by a diverse group of subjects (ages 5–79). Users interacted with four character types: a control (random/idle) model, the emotivector-based model, and standard reactive models found in commercial games like Fable or Oblivion.
Results and Discussion
The study revealed several key insights:
- Task Performance: All subjects completed the game when paired with the emotivector-based character, whereas none finished with the random control.
- Believability Drivers: Statistical analysis indicated that believability was not tied to animation smoothness but rather to the user's ability to perceive the character's attention and emotional state.
- Emergent Behavior: Complex behaviors, such as a "casual look around" versus a "startled glance," emerged naturally from the salience calculations without explicit programming.
- The "Cheating" Effect: The emotivector-based character communicated intent so effectively that some users perceived the character as "cheating" by finding words for them, demonstrating a high level of perceived agency.
Conclusion
The emotivector provides a mathematically grounded, lightweight method for generating believable behavior through low-level anticipation. While currently limited to one-dimensional sensors, it offers a promising complement to higher-level cognitive architectures, grounding abstract reasoning in immediate, sensor-level "feelings."
Summary updated