Fast Neural-Network Approximation of Active Target Search Under Uncertainty

A team from the Technical University of Cluj-Napoca (Yousuf, Lendek, Busoniu) has developed a convolutional neural network (CNN) that approximates the decisions of Active Search (AS) and Intermittent Active Search (ASI) planners for mobile agents searching for stationary targets. The AS and ASI planners use a probability hypothesis density (PHD) filter to estimate the expected number of targets under measurement uncertainty, but require costly online optimization. The CNN replaces this optimization with direct inference, using a multi-channel grid that encodes target beliefs, agent position, visitation history, and boundary information. The network is trained offline on data generated by the AS/ASI planners.

In simulations with both uniform and clustered target distributions, the CNN achieved detection rates comparable to the exact AS/ASI methods while reducing computation time by orders of magnitude. This makes the approach suitable for real-time applications where computational resources are limited, such as drone-based search and rescue, environmental monitoring, or robotic exploration. The work demonstrates that deep learning can effectively approximate complex, uncertainty-aware planning policies, opening the door to faster, more practical autonomous search systems.