Adapting virtual agent interaction style with reinforcement learning to enhance affective engagement

Area of Science:

• Artificial Intelligence
• Human-Computer Interaction
• Affective Computing

Background:

• Dynamic adaptation of communication style is crucial for sustained engagement in human-agent interaction.
• Personalized and engaging dialogue is critical for applications like Behavior Change Interventions.

Purpose of the Study:

• To introduce a reinforcement learning (RL) framework for real-time modulation of agent communication style.
• To maximize the affective valence of the user's emotional response during interaction.
• To create a domain-independent approach for personalized dialogue systems.

Main Methods:

• A between-subjects user study with 20 participants was conducted.
• An adaptive speech-based agent used RL (Thompson Sampling) to select between 'enthusiastic' and 'neutral' styles based on facial emotion recognition.
• The agent adapted its style to increase or maintain user affective valence across interaction turns.

Main Results:

• The RL system successfully adapted its communication style based on user emotional feedback.
• A significant positive correlation was found between users' Psychoticism scores and the reinforcement of the 'neutral' style (Spearman's ρ = 0.70, p = 0.04).
• The adaptive agent maintained usability and personalized interaction based on affective cues, despite no significant differences in user-reported experience compared to a static interface.

Conclusions:

• Adaptive agents hold significant potential for personalizing interaction strategies in emotionally relevant contexts.
• The ability to align agent behavior with user personality profiles supports deployment in long-term interventions.
• Personalized communication enhances user motivation and engagement in therapeutic and behavioral change applications.