Virtual humans making first contact : teaching socially appropriate approaching behavior using deep reinforcement learning

Socially appropriate behavior of humans in public places is governed by various factors and rules. While these aspects have been previously well-explored for virtual humans engaged in focused interactions (e.g. conversations), the same is not true for unfocused interactions (e.g. merely being present in the same social situation), and for the shift from unfocused to focused interactions (e.g. approaching someone for a conversation). The goal of this thesis is to teach virtual humans a sense of appropriate social behavior when approaching a stranger for a conversation in a public place, using deep reinforcement learning. To that end, a model of a socially appropriate approaching behavior is devised, based on empirical studies and previous research. Using deep reinforcement learning in Unity and experimenting with different reward functions, virtual humans (agents) are trained to approach another virtual human (target), while adhering to this model as much as possible. The results are assessed numerically, as well as visually. Finally, a perceptual study compares the best-performing well-trained agent with a baseline agent. The results show that statistically, the well-trained agent causes significantly fewer social offenses (i.e. violates the rules of the model less often) than the baseline agent. The results of the perceptual study also show that participants find the behavior of the well-trained agent overall significantly more socially appropriate than the behavior of the baseline agent. In conclusion, using deep reinforcement learning to teach an agent appropriate social behavior could prove to be a valid alternative to crafting and implementing rule-based behaviors.

Author dc:creator

• Anna Franziska Horne 1989-

Contributors dc:contributor

• Háskólinn í Reykjavík

• Tölvunarfræði
• Meistaraprófsritgerðir
• Gervigreind
• Sýndarveruleiki
• Hermilíkön
• Vélrænt nám
• Samskipti
• Félagsfærni
• Computer science
• Virtual humans (Artificial intelligence)
• Reinforcement learning
• Deep learning (Machine learning)
• Human-computer interaction
• Social interaction
• Social skills

Language dc:language.iso

en

Handle dc:identifier.uri

http://hdl.handle.net/1946/44917

OAI identifier oai:identifier

oai:skemman.is:1946/44917