The extraction of tactile information from an environment depends on the action, sensor design, and the environment itself. This makes identifying an optimal action for a given task, i.e. classification, challenging and typically requires extensive data-rich experiments. We propose utilizing Mutual Information (MI), a rapid-to-evaluate information content met-ric that considers the shared information between two random variables as a means of comparing 'tactile images' from dif-ferent action-sensor-environment pairings. We propose this as a heuristic for selecting actions that offer the highest reliability or the greatest ability to classify different environments. As a demosntration, MI was used to search for an optimal action to distinguish two environments using Bayesian Optimization. The use of MI could guide the data-efficient evaluation and optimization of action selection and multi-modal sensor design.