State Definition in the Tetris Task: Designing a Hybrid Model of Cognition

In defining state/action pairs for reinforcement learning of the Tetris task, we seek to recognize known game states, as well as to learn new ones according to relevant game features. We propose a model of cognition that uses categorization as a mechanism to sort such features into appropriate state-types, and an attention mechanism based on predicted values of each state as a method for deciding which states/features are relevant.