Multi-level metacognition for adaptive behavior

Abstract

Behavior adaptation is an integral aspect for autonomous agents to survive in a world where change is normal. Animals change their foraging routines and socializing habits based on predator risks in their environment. Humans adapt their behavior based on current interests, social norms, stress level, health conditions, upcoming deadlines and various other factors. Artificial agents need to effectively adapt to changes in their environment such that they can quickly adjust their behavior to maintain performance in the changed environment. In this paper, we present a multi-level metacognitive model that allows agents to adapt their behavior in various ways based on the resources available for metacognitive processing. As the agent operates at higher levels of this model, the agent is better equipped to adapt to a wider range of changes. The model has been tested on 2 different applications: (i) a reinforcement learner-based agent trying to navigate and collect rewards in a seasonal grid-world environment and (ii) a convolutional neural network-based agent trying to classify the signals in a radio frequency spectrum world and separate them into known modulations and unknown modulations.