Stopping criteria for ending autonomous, single detector radiological source searches

Area of Science:

• Nuclear Engineering
• Computer Science
• Robotics

Background:

• Traditional radiological source localization relies on statistical algorithms.
• Machine learning, particularly deep and reinforcement learning, has advanced autonomous mobile systems for surveying.
• Existing methods often require operator intervention or perfect source knowledge to terminate searches.

Purpose of the Study:

• To investigate two novel stopping criteria for machine learning-guided radiological source localization systems.
• To compare the performance of a Bayesian/maximum likelihood estimation (MLE) based stopping condition against a machine learning network with a self-stopping action.

Main Methods:

• A convolutional neural network was trained using reinforcement learning in a simulated 10m x 10m environment.
• Two stopping criteria were evaluated: Bayesian likelihood estimation and a dedicated "stop search" action within the neural network.
• Localization accuracy and speed were compared over multiple trials for both methods.

Main Results:

• The Bayesian/MLE approach resulted in a median localization error of ~1.41m and a median speed of 12 steps.
• The machine learning self-stopping action achieved a median localization error of 0m and a median speed of 17 steps.
• The machine learning approach demonstrated superior accuracy in source localization.

Conclusions:

• Machine learning offers effective stopping criteria for autonomous radiological source localization.
• A self-stopping action integrated into the navigation network provides highly accurate, zero-error localization.
• This research presents viable machine learning strategies to improve the efficiency and autonomy of radiological source search systems.