Learning executable agent behaviors from observation
Andrew Guillory, Hai Nguyen, Tucker R. Balch, Charles Lee Isbell · 2006
We present a method for learning a human understandable, executable model of an agent's behavior using observations of its interaction with the environment. By executable we mean that the model is suitable for direct execution by an agent. Traditional models of behavior used for recognition tasks (e.g., Hidden Markov Models) are insufficent for this problem because they cannot respond to input from the environment. We train an Input/Output Hidden Markov Model where the output distributions are mixtures of learned low level actions and the transition distributions are conditional on features detected by the agent's sensors. We show that the system is able to learn both the behavior and human-understandable structure of a simulated model.