Unsupervised Behavior Inference From Human Action Sequences

To effectively coordinate with human teammates in environments where direct communication is limited or unavailable, autonomous agents must be able to infer macro-level strategies from demonstrated micro-level actions in real time. This article introduces UNsupervised Behavior Inference from Action Sequences (UNBIAS): a modular framework for the unsupervised learning of behaviors from sequences of demonstrated human actions. UNBIAS relies on an unsupervised sequential encoder to learn representations of behaviors from sequences of human actions. Behavior representations are then clustered to obtain an unsupervised behavior classification. Real-time strategy prediction is performed using UNBIAS by training a classification layer to learn strategies’ decision boundaries in the behavior representation space and predict cluster labels from encoded incomplete action sequences. In extensive numerical experiments on two real-world datasets of human decision sequences, behavior classifications learned through the UNBIAS framework are shown to better match demonstrator identity groupings than partitions obtained with related unsupervised clustering algorithms, and high-quality online strategy inference is demonstrated. Thus, UNBIAS effectively enables the unsupervised learning of human behaviors from observed actions, enabling autonomous agents to better adapt to human teammates’ strategies in human-machine teams.