Decision Making and Learning for Hybrid Dynamical Agents

Models of human decision making and learning may benefit from being implemented in cognitively-inspired, dynamic frameworks. This paper describes one such framework built on a hybrid dynamical system, which models agent cognition as the continuous evolution of cognitive elements with occasional discrete changes in patterns of behavior. Cognitive elements are beliefs, desires, and intentions, from the Belief-Desire-Intention theory of human rationality, along with ground concepts and sequencing intentions. Continuous dynamics of the hybrid system result from changes in element activation levels; activation can spread to related elements over links in a neurologically-inspired spreading activation network. This hybrid system performs decision making by selecting new patterns of behavior when activation levels of elements reach certain configurations, and performs learning by strengthening spreading activation links between elements. This system is capable of representing long-term sequences of actions, or plans, as well as dynamically replanning by selecting new actions when environmental changes make old action sequences undesirable. In simulations of the model, agents navigated grid world environments while dynamically selecting goals and actions to fulfill those goals, and autonomously learned associations between people and places in their environments.