Deciding where to go is one of the primary challenges in designing an agent that can explore an unknown environment. Grid-worlds provide a flexible framework for representing different variations of this problem, allowing for various types of goals and constraints. Typically, agents move one cell at a time, gathering new information at each time step. However, recomputing a new action after each step can lead to unintended behaviors, such as indecision and forgetting about previous goals. To mitigate this, we define a set of persistent feature layers that can be used by either a linear weighted policy or a neural network approach to identify potential destination locations. The outputs of these policies are processed using knowledge of the environment to ensure that objectives are met in a timely and effective manner. We demonstrate how to train and evaluate a U-Net model in a custom grid-world environment and provide guidance and suggestions for how to use this approach to build complex agent behaviors.