Investigating the Interpretability of ML-Guided Radiological Source Searches

The coupling of reinforcement learning (RL) and deep neural networks (DNN) has demonstrated promising results in many task-oriented scenarios, including radiological source localization. However, these black box approaches present an issue from the user’s perspective – the non-interpretably of results both during and after task completion. In this work, an RL-based convolutional neural network (CNN) for single-detector, radiological source localization is augmented with a system-feedback strategy which provides users with real-time information on estimated search progress, confidence in search termination, and a projection of future actions. In this Q-learning network, the agent is a single, mobile detector; the environment is a 2D simulated space, including background, attenuating obstacles, and source; the actions are (from a top-down view) {left, right, up, down, and stop search}. At each step, the agent moves and records a gamma-ray measurement, the search maps are updated, and the CNN is activated, yielding the action with the greatest q-value (reward). In addition, at each step, system-feedback is generated by virtually probing the network at all locations for q-values. The system-feedback sets are: 1) the confidence in taking each action at the given location, 2) a map of future movements, and 3) a map of stopping likelihood. The information these three sets provide helps the user better understand why a given action was taken and what to expect going forward. The combination of confidence measures, count rate, and pathing does provide interpretable information for discerning current and future actions.