Learn to Behave! Rapid Training of Behavior Automata

摘要

Programming robot or virtual agent behaviors can be a challenging task, and makes attractive the prospect of automatically learning the behaviors from the actions of a human demonstrator. However, learning complex behaviors rapidly from a demonstrator may be difficult if they demand a large number of training samples. We describe an architecture for rapid learning of recurrent behaviors from demonstration. The architecture is based on deterministic hierarchical finitestate automata (HFAs) with classification algorithms taking the place of the state transition function. This architecture allows for task decomposition, statefulness, parameterized features and behaviors, per-behavior feature set customization, and storage of learned behaviors in libraries to be used later on as elements in more complex behaviors. We describe the system, then illustrate its application in a simple, but nontrivial, foraging task involving multiple behaviors.